Changelog

All notable changes to this project will be documented in this file. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Latest 1.0.0Released 2026-07-1520 entries total

1.0.02026-07-15

Added12 entries

Goat is here: a company brain that files what matters from your team's tools and conversations, links it to what you already know, and serves it to any agent over MCP (#554) — @louis.
Goat Brain is a typed, folder-organized knowledge base with an evidence taxonomy, wiki-style links rendered as chips, and first-class binary documents including PDFs (#558, #607, #611, #615, #620, #613) — @louis.
Brains ingest from your sources automatically — Jamie meeting notes, GitHub (PRs, issues, comments), Gmail, Linear, Google Drive, X, and chat captures — filing and linking what matters (#575, #652, #659, #641, #715, #573, #631) — @louis.
Ingestion can enrich entries with web search, respects per-brain spend budgets, and imports company context during onboarding (#691, #725, #724, #728) — @louis.
Goat chat streams live with disconnect-safe, resumable sessions, file attachments that flow end-to-end into the brain, chat pinning, background chats, and Cmd+K to start a new one (#633, #700, #731, #577, #627) — @louis.
Codex-backed sessions run in Goat across cloud and local sandboxes, with in-chat controls, sandbox status, and per-task engine selection (#555, #649, #624, #662, #661) — @louis.
Tasks can run on a schedule, notify their originating chat on completion, expose their harness config, and sit behind a background-task preference (#560, #557, #559, #717) — @louis.
A single user-level MCP endpoint at /mcp exposes every brain you can access, with a personal MCP settings tab and member onboarding (#739, #600, #714) — @louis.
Workspaces gate brain access with member-level permissions, member-contributed sources, multi-account connections, and cross-member deduplication (#596, #656, #740) — @louis.
Workspace billing ships with pooled monthly allowances, per-seat pricing, USD credits, per-brain intelligence tiers, credit top-up promotion codes, and a spend overview with charts (#718, #732, #742, #745, #648, #651) — @louis.
A brain overview dashboard, activity feed, and settings overview give each brain a home, alongside a model picker, Fumadocs-powered docs, browser research tools, an onboarding flow, and new app icons (#741, #625, #669, #594, #572, #576, #710, #727) — @louis.
Codex chats get auto-generated titles, and Goat gained an appearance theme toggle and a quicker new-brain action in the sidebar (#668, #703, #702) — @opencompany-bot.

Changed4 entries

Brain retrieval was rebuilt on a Postgres-native read plane and unified into a single surface; chat and MCP now read the brain read-only, with all writes flowing through ingestion (#621, #695, #690) — @louis.
Navigation across the brain and app is now client-first for instant transitions (#591, #635) — @louis.
The sidebar gained a workspace switcher in its header, recent-chat history, archive-on-hover, and a per-brain context menu (#667, #696, #698, #712) — @louis.
The ingestion model was updated, the per-event ingestion budget was raised to one dollar, and the browser tab now reads "OpenCompany" (#655, #734, #747) — @louis.

Fixed4 entries

Brain ingestion no longer suffers retry storms or conflict-page pollution, and retrieval tracking is accurate (#707, #743) — @louis.
Fixed an infinite render loop that froze brain navigation and a composer freeze when input scrolled past its max height (#736, #730) — @louis.
Goat chat session continuity and post-completion navigation are reliable, and chat pin transactions are consistent (#590, #706, #737) — @louis.
Jamie webhook API-key handling and setup were corrected (#580, #654) — @louis.

0.18.02026-06-25

Added9 entries

Agents can now run Codex-backed sessions with persisted auth, sandbox lifecycle support, runtime event ingestion, and transcript UI support behind the Codex engine flag (#525) — @louis.
Google Drive is now available as a first-party integration with OAuth, Drive search/read/export tools, and Google Docs create/update tools (#515) — @louis.
Company workspaces now have a dedicated Integrations tab for first-party and MCP integration controls (#522) — @louis.
Personal routines now include scheduling with create, edit, delete, toggle, run-now, next-run, and timezone controls (#458, #507) — @louis.
Billing now supports weekly spending limits and automatic credit refills with saved cards (#505) — @jasper.
Delegated coding and memory tools now use a runner-hosted LLM token broker, keeping raw provider keys out of sandboxes while metering brokered usage server-side (#439) — @louis.
Z.ai GLM-5.2 is now available in the model catalog with billing and reasoning-capable model support (#486) — @opencompany-bot.
Local development can now persist Turbo logs and inspect them through bun run dev:logs (#518) — @louis.
Codex E2B toolbox templates now include Playwright and Playwright-managed Chromium for browser-ready coding sessions (#525) — @louis.

Changed6 entries

Mobile Memory pages now let users swipe the file tree away for a full-width file view, while guarding iOS edge-back gestures (#510) — @jasper.
Tool-call cards now show the brand icon for external services like Linear, Slack, Gmail, GitHub, Notion, PostHog, Google Calendar, Google Drive, and Neon (#524) — @jasper.
After-session memory updates now appear as a compact footer badge with a tooltip and link to the spawned memory-pass session (#526) — @louis.
The personal and company sidebar space switcher is now a dropdown, and the private beta badge has been removed (#520) — @louis.
Feedback reports now use a one-tap Bug, Feedback, and Idea segmented selector with Feedback selected by default (#499) — @jasper.
Company and personal route navigation now uses faster links, debounced prefetching, and loading skeletons for slow routes (#523) — @louis.

Fixed9 entries

Google Drive OAuth now uses the narrower drive.file scope so production connections are not blocked by full-Drive restricted-scope verification (#517) — @louis.
iOS PWA layouts now respect safe-area insets, avoid keyboard popups in the mobile model picker, use full-width chat messages on phones, and prevent mobile settings clipping (#501, #504) — @jasper.
MCP reconnect and tool-discovery failures now return recoverable tool outputs instead of failing the whole session (#503) — @louis.
Durable Stream publishing now retries with a fresh producer when a cached producer has already closed, improving live delivery for lifecycle and memory events (#502) — @louis.
The automatic-refill Add a card flow now passes the required Stripe Checkout currency (#511) — @jasper.
Personal session unread dots are visible again and are no longer re-armed by background after-session memory passes (#509, #513) — @jasper.
Existing personal agents now backfill missing soul.md files so the Soul nav opens the editor instead of jumping to a new chat (#514) — @jasper.
Company sidebars no longer show personal-agent threads in recent or pinned session history (#521) — @louis.
Reading position is preserved when completed work blocks collapse or expand in long session transcripts (#508) — @jasper.

0.17.02026-06-22

Fixed1 entry

Completed assistant replies now appear live when the final stream payload is missed, without needing a page reload (#496) — @louis.

0.16.02026-06-19

Added2 entries

Personal sessions on mobile can now swipe from the opposite edge for session details, while filmstrip gestures respect swipe direction (#492, #495) — @jasper.
The personal agent sidebar now includes a tabbed session details inspector (#489) — @louis.

Changed4 entries

Feedback reports now submit optimistically and archive actions show a confirmation toast (#491) — @jasper.
Personal Brain and after-session memory guidance is stricter, so agents only save persistent files or memories when user intent is clear (#485) — @louis.
The personal sidebar home entry is now labeled New Session and uses the send icon (#487) — @louis.
Personal sidebar session lists now hide scrollbars while remaining scrollable (#494) — @jasper.

Fixed2 entries

Agent replies no longer reorder in the transcript when stream events arrive out of order (#493) — @jasper.
Mobile drawer swipes now work more reliably in Safari while preserving normal vertical page scrolling (#490) — @jasper.

0.15.02026-06-18

Added2 entries

Personal Brain now keeps restorable file versions, with restore tooling, caps, and runner guidance for safer durable writes (#483) — @jasper.
Mobile users can now swipe to open the main navigation drawer (#484) — @jasper.

Fixed4 entries

The onboarding wizard no longer resets when setting tool policies during the capabilities step (#481) — @jasper.
The three-dot menu in the brain file view no longer renders behind the file text (#480) — @jasper.
GitHub work repositories granted after connecting now resolve via a live lookup instead of being missed (#479) — @jasper.
Various mobile and phone UI fixes (#478) — @jasper.

0.14.02026-06-16

Added5 entries

OpenRouter Fusion models are now supported (#473) — @louis.
Messages sent mid-run now offer Steer, Queue, and Interrupt send-modes (#468) — @jasper.
Chat sessions can now be renamed via double-click or a right-click menu (#467) — @jasper.
Clarifying questions now always include a free-text "Other" answer (#466) — @jasper.
Screenshots dropped into chat are now attached to the Linear issue created from that conversation (#464) — @jasper.

Changed3 entries

Built-in skills now show distinct icons instead of a shared Sparkles placeholder (#475) — @jasper.
Memory lookup prompting tightened for more relevant recall (#474) — @louis.
Typed @mentions in the agent editor are no longer auto-converted as you type (#470) — @jasper.

Fixed4 entries

Near-identical person records in memory are now deduplicated (#477) — @jasper.
The thinking trace now separates distinct reasoning phases instead of running them together (#476) — @jasper.
The "Updated memory" card no longer overlaps the message footer (#469) — @jasper.
Fixed the personal billing redirect (#465) — @louis.

0.13.02026-06-12

Added8 entries

Agents can now create Linear issues directly from chat, preserving the relevant conversation context in the issue (#462) — @jasper.
Notion is now available as a first-party MCP integration with OAuth connection flow and tool discovery (#455) — @louis.
New users can now complete a Leo-led onboarding flow with updated personal onboarding screens and kickoff behavior (#451) — @louis.
Agents can now use the bundled YC knowledge skill (#456) — @louis.
Feedback reports can now include screenshots, which are uploaded and embedded inline in the linked Linear issue (#447) — @jasper.
The personal Memory surface is now URL-addressable, so selected memory paths can be opened and restored from direct links (#446) — @jasper.
Memory queries now support relative --since windows and query-less recent-memory listings (#444) — @louis.
Preview workspaces now show an early-preview badge in the sidebar (#457) — @louis.

Changed3 entries

Personal sidebar navigation was refreshed for the updated personal surface structure (#449) — @louis.
GitHub permission scopes were consolidated and renamed around clearer permission groups (#450) — @louis.
Signup welcome email delivery and unsubscribe handling were refactored for more reliable account email flows (#452) — @louis.

Fixed4 entries

Home-space image drag and drop now preserves sent images and shows them in the chat bubble after sending (#454) — @jasper.
Sessions now recover stale stream state after returning to a tab instead of staying stuck on outdated loading state (#448) — @louis.
Personal Brain writes are no longer silently dropped when the agent bundle exceeds the 80-file cap (#445) — @louis.
Chat scrolling now settles at the bottom when expected during active sessions (#431) — @jasper.

0.12.02026-06-11

Added10 entries

Brain files created or edited by agents now show as clickable attachments under assistant turns, and the company Brain can be opened directly at /brain/<path> deep links (#356) — @louis.
The personal Brain is now URL-addressable too, with nested file links restoring the selected file on reload and keeping the Brain navigation active (#426) — @jasper.
Personal Brain files created or edited during agent turns now link directly from the session transcript into the matching personal Brain path (#442) — @louis.
Personal sessions now support split-screen panes, so multiple chats can stay visible and be arranged side by side (#441) — @louis.
Agents can now start scoped subagents with selected tool grants, with subagent lifecycle events shown in the parent session transcript (#421) — @louis.
The personal Integrations tab now expands inline with connected account details, permission summaries, degraded-access warnings, disconnect actions, and GitHub repository refresh (#390) — @louis.
Personal recall can now be constrained by time windows, including query-less recent-session snippets, so agents can ask for memory from a specific lookback period (#416) — @louis.
Large pasted text is captured through the attachment pipeline, and oversized text attachments are materialized as sandbox file references instead of being inlined into every model call (#410) — @louis.
Session usage now updates live during streaming, long-running tool calls show an elapsed counter, and the context-window tooltip includes session cost (#395, #376, #399) — @louis.
Runner observability now records per-tool-call phase timings, Braintrust spans, PostHog rollups, and hidden sandbox hydration timing events for production latency diagnosis (#411, #417) — @louis.

Changed10 entries

Starting a session from the personal home prompt now switches to an optimistic session view immediately while creation and navigation finish in the background (#420) — @louis.
Runner sandbox hydration now overlaps bundle, skill, and setup work to reduce startup latency for sessions that need a sandbox (#434) — @louis.
Personal inbox cards now render markdown in their body and step text, matching the markdown contract used by the inbox tool (#438) — @louis.
Recall recap guidance is tighter, with cleaner formatting for recalled memories and follow-up summaries (#436) — @louis.
Preview deployment PR comments now include ?skipOnboarding, letting reviewers test fresh preview accounts without going through onboarding (#435) — @louis.
New personal agents now default to Leo on Kimi K2.6, with updated default capabilities, model selection, settings reset, and personal agent source normalization (#409, #425) — @louis.
Memory tooling now returns richer memory query and memory get output, clearer invalid-usage help, a tighter keeper kickoff prompt, and canonical guidance that durable facts belong in structured memory while agent/user.md is the always-loaded digest (#413, #418, #422, #428) — @louis.
Memory keeper passes now use Gemini 3.1 Flash Lite, surface what they stored, and appear in the transcript near the turn they followed instead of pinned to the bottom (#401, #397) — @louis.
Runtime details now stay collapsed by default for new sessions, including sessions with related parent or child metadata (#423) — @louis.
Slack support channel names now keep the per-workspace suffix at the end, for example <customer>-x-opencompany-<id8> (#394) — @jasper.

Fixed8 entries

Slash command session starts now route to the intended personal or company surface instead of opening the wrong session view (#432) — @louis.
Session lists now consistently sort by creation time across the company and personal sidebars (#437) — @louis.
Personal agent skill slash commands are available again in the composer slash menu and insert the right skill mention (#440) — @louis.
Personal-session inspector links now stay inside the personal surface instead of opening related sessions, session pages, or agent links under /company (#391) — @louis.
Personal runtime tool guidance now advertises the correct work/, personal-brain/, and agent/ roots and blocks generic file tools from editing structured memory paths (#414) — @louis.
OpenCode tool usage is now parsed from current step_finish events and charged back to the spawning session instead of silently appearing as zero-cost work (#402) — @jasper.
Runner shutdown and interruption handling now drains in-flight runs, persists abort/failure state more reliably, and avoids replaying interrupted deploy-time work (#378) — @louis.
E2B stream callbacks are guarded against detached rejections so stopping an opencode_coder command no longer crashes the runner process (#400) — @louis.

0.11.02026-06-10

Added6 entries

The new /personal experiment surface gives each user a private default agent with its own sessions, behavior editor, capability panels, context files, memory, tools, skills, integrations, and settings views (#353) — @louis.
Claude Fable 5 is now available end to end in the model catalog, with reasoning, image, PDF, long-context, and billing support (#384) — @louis.
Preview deployments can now complete Gmail and Google Calendar OAuth through a stable callback broker, making Google integration testing work across hosted preview URLs (#382) — @louis.
Public changelog entries can now include embedded images and short looping videos, with docs for recording, compressing, uploading, and referencing changelog media (#387) — @louis.
The account menu now shows when the currently running build was last deployed, with release details in the tooltip (#388) — @louis.
Personal agents now support plain @github for all installation repositories and repo-scoped @github/owner/repo mentions, with matching sandbox and coding-tool access (#392) — @louis.

Changed1 entry

Sandbox preparation no longer waits on optional developer-tool installation during the critical setup path (#393) — @louis.

Fixed1 entry

Personal recall searches no longer fail before running when the runner configures its statement timeout (#389) — @louis.

0.10.02026-06-08

Added3 entries

Paste, drag, drop, or pick screenshots, PDFs, and text/code files in the session composer; attachments are validated per model, stored in private Vercel Blob, rendered as compact cards, and inlined into runner model calls (#362) — @jasper.
Attached skills now appear in the composer slash menu as /<command> entries, using the skill's declared command frontmatter when present and inserting the matching @skill/<id> mention when selected (#350) — @louis.
Brain pages now show when the file was last updated directly in the header, and include a copy button for the page title plus current contents (#355, #358) — @louis, @jasper.

Changed2 entries

Session pages now use the session title as the browser tab title, making multiple open chats easier to tell apart (#354) — @louis.
Slack Connect support channels now use safer <customer>-<id8>-x-opencompany naming, include recovery sweeps for stuck provisioning, and have expanded setup and operations docs (#335) — @jasper.

Fixed3 entries

opencode_coder timeouts and command failures now preserve partial diffs, persist artifacts, return a resumable opencode session id, and report timeout status instead of dropping the work (#344) — @louis.
Long-running tool calls no longer trigger runner job replay and double billing after transient lease gaps; job lease TTL and lease-busy retry ceilings are now configurable (#345) — @louis.
Coding sandboxes now configure GitHub git credentials and best-effort ensure rg and bun are available, so plain git commands, repo search, and tests work more reliably inside sessions (#346) — @louis.

0.9.02026-06-05

Added7 entries

Gmail and Google Calendar are now first-party integrations — connect one or more Google accounts to give agents read-only Gmail and read/write Calendar tools, with per-account calendar selection (#319, #320) — @louis.
Switch the model for a single chat on the fly without changing the agent's saved default, in a redesigned composer with a toolbar tray for the agent and model selectors (#331) — @louis.
A context-window usage ring in the session top bar shows how full the model's context window is for the current turn (#332) — @louis.
Slack Connect — completing onboarding now provisions a private shared Slack channel with our team, surfaced as a "Connect on Slack" card on your workspace home (#278) — @jasper.
Tool calls that touch a Linear issue now show an external-link icon that opens the issue directly (#324) — @jasper.
Deferred tool discovery — capability tools and coding agents are no longer all preloaded; the model discovers them on demand via find_tools and runs them through a use_tool dispatcher, improving tool selection and shrinking the cached prompt (#321) — @louis.
New shared @opencompany/ui component package (shadcn / Tailwind v4) and a design-system app to document and showcase it (#323) — @louis.

Changed7 entries

Deferred-tool arguments now run through a validate → coerce → repair pipeline before dispatch, so smaller models' malformed tool calls are fixed automatically instead of failing (#328) — @louis.
Brain access scope is now enforced at the tool layer, so out-of-scope Brain writes are rejected up front with a clear error instead of being silently dropped at sync (#327) — @louis.
The new-session empty state no longer shows default prompt chips (#315) — @louis.
The /btw start toast no longer echoes the prompt; it shows just a confirmation and an Open action (#330) — @louis.
Faster navigation — removed the route-level session loader that flashed skeletons on fast nav (#303) — @louis.
Production deploys are faster and now recover gracefully from interrupted runner deploys (#310, #318) — @louis.
Runner model and tool calls are now traced via Braintrust's wrapAISDK defaults for richer, lower-maintenance observability (#304) — @louis.

Fixed3 entries

Unknown-tool and invalid tool-input calls are now recovered as tool results that steer the model back on track, instead of crashing the turn (#314, #325) — @louis.
Fixed a startup-status race that briefly showed "Stopped before finishing" on a freshly started session that was actually still running (#306, #312) — @louis.
Runner completions that contain only reasoning and no answer are now rejected instead of recorded as a successful turn (#300) — @louis.

Security1 entry

Hardened GitHub permission gating for agent shell commands, closing permission bypasses (#302) — @louis.

0.8.02026-06-04

Added8 entries

Real-time session sync and durable token streaming powered by Electric and Durable Streams, so chat updates propagate live and survive interrupted connections (#283) — @louis.
New opencode coding agent tool, letting agents run coding tasks through opencode (#266) — @louis.
Official PostHog MCP support with OAuth / Dynamic Client Registration, so agents can connect to PostHog without manual token wiring (#269) — @louis.
E2B sandbox compute is now metered and billed as agent usage (#264) — @louis.
Upload a custom profile picture for your account (#257) — @jasper.
Credit top-up analytics to track balance refills (#279) — @louis.
OpenCompany icon assets (#284) — @louis.
External skills support (V1) — attach a public GitHub repo or skills.sh page to an agent via an @skill mention; the skill is snapshotted workspace-side and materialized read-only into the runner sandbox at session start alongside built-in skills (#296) — @louis.

Changed9 entries

Session detail now paints instantly from local TanStack DB collections instead of waiting on the server (#290) — @louis.
Improved AI SDK prompt caching for lower latency and cost on repeated context (#275, #285) — @louis.
The agent self-edit skill now asks before making changes, stays lightweight, and better equips coding agents (#281) — @louis.
Onboarding call step reworked so its actions fit on a single screen (#263) — @jasper.
Web functions now run in the Frankfurt region for lower latency to European users (#286) — @louis.
MCP server tools are now lazy-loaded via per-server search_tools / use_tool meta-tools, so raw tool schemas are only sent to the model on demand instead of being injected on every turn — eliminating the per-turn token cost of large MCP catalogs like Slack (#298) — @louis.
Starting a new session from the root-route prompt now navigates instantly; session detail is assembled in-memory from the just-inserted rows instead of waiting on seven follow-up database reads (#295) — @louis.
E2B sandbox warm-up now begins at the start of a turn so the ~10 s cold-start overlaps the model's initial token stream rather than blocking the first tool call (#297) — @louis.
Runner parallel-session concurrency is now configurable via RUNNER_WORKER_CONCURRENCY (default 8); production limit raised to 40 (#294) — @louis.

Fixed10 entries

Coalesced durable-stream token appends to fix streaming lag in production (#288) — @louis.
Electric and Durable Streams now work in local development (#291), with producer linger and in-flight limits tuned for throughput (#292) — @louis.
Tool output previews are now redacted of secrets before being persisted to the database (#273) — @louis.
Approval stream state no longer gets stuck out of sync (#272) — @louis.
opencode now resolves the correct attached repository (#271) — @louis.
Linear save_* MCP tools are no longer misclassified as Admin-only (#267) — @louis.
@mention pills no longer break after an agent edits itself (#268) — @louis.
A run is never replayed after its lease expires, and onboarding dedup is hardened against duplicate runs (#262) — @louis.
The Render deploy trigger is now resilient to empty API responses (#287) — @louis.
Session status and error no longer flicker blank at the start of a session; the live stream only takes authority over the Postgres snapshot once it has received a status-bearing event (#299) — @louis.

0.7.02026-06-03

Added5 entries

Agents can now pause mid-run to ask structured questions with single-select, multi-select, and "Other" answers, then continue from the user's response (#255) — @louis.
Onboarding agents now use structured questions and receive signup names in their starting context, making first-run setup more personal and less brittle (#259, #260) — @louis.
Apify-backed Instagram and TikTok tools for public profiles, posts, reels, videos, comments, search, and async social scraping jobs (#253) — @louis.
Searchable, provider-grouped model picker with Capability, Speed, and Cost ratings so models are easier to compare at a glance (#254) — @louis.
Day-one Brain wiki and private agent agent/soul.md scaffolding, giving new workspaces a cleaner shared knowledge structure and better self-configuration defaults (#251) — @louis.

Changed2 entries

Onboarding call booking now uses one booking-aware button, the official Cal.com embed bootstrap, and clearer new-tab labeling (#249, #250, #258) — @jasper.
Public changelog entries now include author attributions (#256) — @louis.

Fixed4 entries

Chat scroll-to-top behavior after sending a message is now stable during streaming and viewport resizing, with less jitter and less fighting user scrolls (#248) — @jasper.
Paused sessions now refresh stream tokens correctly, and settings hydrate consistently after reloads (#252) — @louis.
Editing a markdown title or header no longer flickers while typing (#201) — @jasper.
Naming a new Brain file from its tab now updates the sidebar tree immediately (#198) — @jasper.

0.6.02026-06-02

Added11 entries

Guided first-run setup — after signing up you land in a live setup conversation that helps configure your workspace, instead of an empty editor (#241) — @louis.
New /btw command to start a new session without leaving your current view (#243) — @louis.
New /clear command in the chat composer (#231) — @louis.
"Run now" button to trigger a scheduled agent immediately (#236) — @louis.
Agents can now pull content from TikTok and Instagram (#239) and YouTube (#233) — @louis.
Pin sessions to the top of the sidebar (#213) — @jasper.
Live product status indicator in the sidebar, linking to the public status page (#224) — @louis.
Agents can now adjust their own schedules (#217) — @louis.
Workspace tool permissions — allow, ask, or deny which tools and integrations an agent can use, with approval prompts that are remembered between runs (#229) — @louis.
Live reasoning support for Kimi models (#226) — @louis.
Provider logos for xAI and MiniMax in the model picker (#232) — @jasper.

Changed5 entries

Agent runs no longer break when an integration like Linear or Slack is turned on but not yet connected — the agent now asks you to connect it instead (#238) — @louis.
When you send a message, it scrolls to the top of the chat, matching the ChatGPT/Claude experience (#219) — @jasper.
The copy button now stays visible on assistant replies and appears on hover for your own messages (#222) — @jasper.
The onboarding call step now has a clear "Skip for now" button (#220) — @jasper.
The "reconnecting" banner no longer appears in chat during normal streaming; connection status now lives in the inspector instead (#242) — @louis.

Fixed6 entries

Linear no longer shows as disconnected after you reconnect with valid credentials (#234) — @louis.
Agent turns that quietly stop mid-task are now flagged instead of being recorded as a silent success — important for scheduled and unattended runs (#225) — @louis.
Sessions now recover when a response stream gets interrupted (#223) — @louis.
Rapidly toggling a session's pin no longer causes glitches (#218) — @jasper.
Restored the after-session hook menu when typing at the start of a line (#227) — @jasper.
The hook suggestion popup now stays beneath the model dropdown (#230) — @jasper.

0.5.02026-06-01

Added7 entries

Cron-based agent schedules — agents can now be scheduled to run on a recurring basis (#208) — @louis.
Read-only X (Twitter) hosted tool for agents (#212) — @louis.
Agent self-update skill, allowing agents to evolve their own .agent definition (#210) — @louis.
Dark mode support (#186) — @louis.
Grok model support via the AI gateway (#206) — @louis.
MiniMax and Kimi gateway models (#193) — @louis.
Client-side felt time-to-first-token measurement (#192) — @louis.

Changed5 entries

GitHub repositories are now decoupled from the amp tool, making repository bindings more flexible (#204) — @louis.
Lazy GitHub sandbox initialization is more solid and reliable (#207) — @louis.
Vercel Skew Protection is enabled via custom deploymentId for safer deploys (#205) — @louis.
Onboarding call booking step is now skippable (#191) — @jasper.
Assistant turn duration is now shown next to the copy button (#187) — @jasper.

Fixed6 entries

Agents no longer disappear from the list after creating or editing a new agent (#203) — @jasper.
Renaming an agent no longer resets it to "Untitled agent" (#194) — @jasper.
Vercel deployment ID length is correctly handled (#211) — @louis.
Session archiving is now optimistic and instant instead of waiting on the server (#200) — @jasper.
Feedback dialog auto-closes after a successful submit (#190) — @jasper.
Page no longer rubber-bands at the fold due to overscroll-behavior fix (#189) — @jasper.

0.4.12026-05-30

Changed3 entries

Runner job worker now wakes immediately on enqueue instead of waiting for the next poll interval, reducing time-to-first-token for agent sessions — @louis.
Web dispatch no longer blocks on analytics flush; PostHog capture and runner dispatch now run concurrently — @louis.
Workspace context loading collapses user, workspace, and role into a single joined query, and session submission runs the balance check and auth lookup concurrently, cutting pre-dispatch database round-trips — @louis.

Fixed1 entry

Runner event serialization no longer crashes when created_at arrives as a string from the raw lease-write path; the value is now coerced to a Date at the source — @louis.

0.4.02026-05-28

Added7 entries

Workspace integration resource bindings with encrypted credential storage, multi-connection GitHub repository binding, and runtime validation for bound resources.
Linear and Slack MCP support, including workspace setup flows, agent editor tool gating, runner-side MCP execution, and .agent file configuration.
Delegated agent sessions through @agent/<slug> mentions and a delegate_to_agent tool, with child-session tracking and parent usage rollups.
Repo-scoped GitHub auth for agent shell commands and AMP runs, with ephemeral credentials and output redaction.
GPT 5.2 Codex support and configurable AMP execution modes.
Agent editor support for markdown headings, ordered and unordered lists, and automatic conversion of typed or pasted @/# mentions into pills.
A richer chat composer with quick-start chips, auto-resize, stop controls, copy-message actions, drag-and-drop and paste handling shells, and improved keyboard hints.

Changed3 entries

Local setup now keeps WorkOS authentication redirects on localhost:3000 while continuing to use ngrok for GitHub and other public integration callbacks.
Production web deploys now rely on the GitHub Actions release workflow instead of Vercel Git-triggered deployments.
AMP command handling now avoids rejected stream-JSON mode combinations and falls back to sanitized plain output when structured results are unavailable.

Fixed5 entries

GitHub integration sync in production now avoids unsupported Neon HTTP transactions and normalizes legacy agent configs.
Sandbox repository clones, AMP publishing, and authenticated shell commands now handle GitHub App credentials more reliably.
Session chat no longer allows duplicate submissions while an assistant is responding and no longer leaves failed sessions stuck in a thinking state.
The New agent button now shows immediate pending feedback, blocks double-click creation, and surfaces creation failures.
Linear MCP credential setup now reports missing encryption-key configuration with clearer diagnostics.

0.3.02026-05-27

Added6 entries

Deterministic edit_file support for hosted agents, giving runner sessions a safer and more reviewable way to modify files.
Expanded hosted Exa tools and clearer search schema guidance for agent research workflows.
Common Vercel AI Gateway model presets across the editor, runtime configuration, and billing calculations.
Model and turn-cost analytics for better visibility into session usage and credit spend.
Signup welcome emails, session-aware feedback reports, and an onboarding call booking step.
A pulsing sidebar indicator for active sessions.

Changed4 entries

Runner internals now separate session lifecycle, job leasing, tool dispatch, model streaming, and usage recording for more reliable hosted runs.
Agent file parsing, mention handling, runtime types, and after-session behavior now live in the shared agent runtime package used by both web and runner.
GitHub sync jobs now have retry and sweeper support to recover pending workspace file updates more reliably.
Development and review tooling now includes CodeRabbit configuration and more consistent local script environment loading.

Fixed4 entries

Blank user messages now survive session payload handling and runtime event rendering.
Agent detail mention caches no longer leak across agent boundaries.
Hosted tool fan-out is capped to avoid runaway parallel tool execution.
GitHub integration callback diagnostics now expose enough detail to troubleshoot failed setup flows.

0.2.02026-05-26

Added7 entries

Git-backed agent editing, including file-backed .agent sync, rename propagation, and visible GitHub sync state.
Hosted agent sessions with root-route session start, run controls, archive lifecycle, automatic sandbox pausing, and first-message session titles.
Agent tools for hosted Exa search and lightweight web fetching.
WorkOS organization workspace support, signup onboarding, default workspace credits, and personal environment setup.
Brain context files for agents, plus a more complete workspace file experience.
Session usage, billing, analytics, and cached-token tracking so teams can understand credit spend.
Public docs, public changelog, and in-app feedback intake.

Changed5 entries

Workspace sessions now load from cache first and route transitions respond faster.
Session UI now better explains reasoning model choice, token usage, tool replay, and follow-up state.
Workspace loading and agent navigation states are more predictable.
Release automation now coalesces CI, syncs Inngest production functions, records GitHub deployments, and runs on Blacksmith.
Observability now covers agent sync, launch events, runner logs, and Better Stack error context.

Fixed5 entries

AuthKit sign-in and sign-up redirects now use full document navigation when required.
Runner sandboxes recover more reliably from E2B lifecycle and tool failures.
Brain new file and new folder actions no longer throw on click.
Direct session title generation and tool replay behave consistently.
Production release smoke checks and Render release waiting are more reliable.

Removed1 entry

Persisted tool delta events and the sidebar running badge to reduce noisy state.

0.1.02026-05-20

Added3 entries

Initial Next.js app prototype with sidebar navigation, Agents, Brain, Inbox, Settings, and session surfaces.
Early WorkOS authentication, branchable Neon database setup, and persistent agents.
Turborepo workspace structure with Bun, shared database package, CI, formatting, tests, and secret scanning.

Changelog

All notable changes to this project will be documented in this file. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Latest 1.0.0Released 2026-07-1520 entries total

1.0.02026-07-15

Added12 entries

Goat is here: a company brain that files what matters from your team's tools and conversations, links it to what you already know, and serves it to any agent over MCP (#554) — @louis.
Goat Brain is a typed, folder-organized knowledge base with an evidence taxonomy, wiki-style links rendered as chips, and first-class binary documents including PDFs (#558, #607, #611, #615, #620, #613) — @louis.
Brains ingest from your sources automatically — Jamie meeting notes, GitHub (PRs, issues, comments), Gmail, Linear, Google Drive, X, and chat captures — filing and linking what matters (#575, #652, #659, #641, #715, #573, #631) — @louis.
Ingestion can enrich entries with web search, respects per-brain spend budgets, and imports company context during onboarding (#691, #725, #724, #728) — @louis.
Goat chat streams live with disconnect-safe, resumable sessions, file attachments that flow end-to-end into the brain, chat pinning, background chats, and Cmd+K to start a new one (#633, #700, #731, #577, #627) — @louis.
Codex-backed sessions run in Goat across cloud and local sandboxes, with in-chat controls, sandbox status, and per-task engine selection (#555, #649, #624, #662, #661) — @louis.
Tasks can run on a schedule, notify their originating chat on completion, expose their harness config, and sit behind a background-task preference (#560, #557, #559, #717) — @louis.
A single user-level MCP endpoint at /mcp exposes every brain you can access, with a personal MCP settings tab and member onboarding (#739, #600, #714) — @louis.
Workspaces gate brain access with member-level permissions, member-contributed sources, multi-account connections, and cross-member deduplication (#596, #656, #740) — @louis.
Workspace billing ships with pooled monthly allowances, per-seat pricing, USD credits, per-brain intelligence tiers, credit top-up promotion codes, and a spend overview with charts (#718, #732, #742, #745, #648, #651) — @louis.
A brain overview dashboard, activity feed, and settings overview give each brain a home, alongside a model picker, Fumadocs-powered docs, browser research tools, an onboarding flow, and new app icons (#741, #625, #669, #594, #572, #576, #710, #727) — @louis.
Codex chats get auto-generated titles, and Goat gained an appearance theme toggle and a quicker new-brain action in the sidebar (#668, #703, #702) — @opencompany-bot.

Changed4 entries

Brain retrieval was rebuilt on a Postgres-native read plane and unified into a single surface; chat and MCP now read the brain read-only, with all writes flowing through ingestion (#621, #695, #690) — @louis.
Navigation across the brain and app is now client-first for instant transitions (#591, #635) — @louis.
The sidebar gained a workspace switcher in its header, recent-chat history, archive-on-hover, and a per-brain context menu (#667, #696, #698, #712) — @louis.
The ingestion model was updated, the per-event ingestion budget was raised to one dollar, and the browser tab now reads "OpenCompany" (#655, #734, #747) — @louis.

Fixed4 entries

Brain ingestion no longer suffers retry storms or conflict-page pollution, and retrieval tracking is accurate (#707, #743) — @louis.
Fixed an infinite render loop that froze brain navigation and a composer freeze when input scrolled past its max height (#736, #730) — @louis.
Goat chat session continuity and post-completion navigation are reliable, and chat pin transactions are consistent (#590, #706, #737) — @louis.
Jamie webhook API-key handling and setup were corrected (#580, #654) — @louis.

0.18.02026-06-25

Added9 entries

Agents can now run Codex-backed sessions with persisted auth, sandbox lifecycle support, runtime event ingestion, and transcript UI support behind the Codex engine flag (#525) — @louis.
Google Drive is now available as a first-party integration with OAuth, Drive search/read/export tools, and Google Docs create/update tools (#515) — @louis.
Company workspaces now have a dedicated Integrations tab for first-party and MCP integration controls (#522) — @louis.
Personal routines now include scheduling with create, edit, delete, toggle, run-now, next-run, and timezone controls (#458, #507) — @louis.
Billing now supports weekly spending limits and automatic credit refills with saved cards (#505) — @jasper.
Delegated coding and memory tools now use a runner-hosted LLM token broker, keeping raw provider keys out of sandboxes while metering brokered usage server-side (#439) — @louis.
Z.ai GLM-5.2 is now available in the model catalog with billing and reasoning-capable model support (#486) — @opencompany-bot.
Local development can now persist Turbo logs and inspect them through bun run dev:logs (#518) — @louis.
Codex E2B toolbox templates now include Playwright and Playwright-managed Chromium for browser-ready coding sessions (#525) — @louis.

Changed6 entries

Mobile Memory pages now let users swipe the file tree away for a full-width file view, while guarding iOS edge-back gestures (#510) — @jasper.
Tool-call cards now show the brand icon for external services like Linear, Slack, Gmail, GitHub, Notion, PostHog, Google Calendar, Google Drive, and Neon (#524) — @jasper.
After-session memory updates now appear as a compact footer badge with a tooltip and link to the spawned memory-pass session (#526) — @louis.
The personal and company sidebar space switcher is now a dropdown, and the private beta badge has been removed (#520) — @louis.
Feedback reports now use a one-tap Bug, Feedback, and Idea segmented selector with Feedback selected by default (#499) — @jasper.
Company and personal route navigation now uses faster links, debounced prefetching, and loading skeletons for slow routes (#523) — @louis.

Fixed9 entries

Google Drive OAuth now uses the narrower drive.file scope so production connections are not blocked by full-Drive restricted-scope verification (#517) — @louis.
iOS PWA layouts now respect safe-area insets, avoid keyboard popups in the mobile model picker, use full-width chat messages on phones, and prevent mobile settings clipping (#501, #504) — @jasper.
MCP reconnect and tool-discovery failures now return recoverable tool outputs instead of failing the whole session (#503) — @louis.
Durable Stream publishing now retries with a fresh producer when a cached producer has already closed, improving live delivery for lifecycle and memory events (#502) — @louis.
The automatic-refill Add a card flow now passes the required Stripe Checkout currency (#511) — @jasper.
Personal session unread dots are visible again and are no longer re-armed by background after-session memory passes (#509, #513) — @jasper.
Existing personal agents now backfill missing soul.md files so the Soul nav opens the editor instead of jumping to a new chat (#514) — @jasper.
Company sidebars no longer show personal-agent threads in recent or pinned session history (#521) — @louis.
Reading position is preserved when completed work blocks collapse or expand in long session transcripts (#508) — @jasper.

0.17.02026-06-22

Fixed1 entry

Completed assistant replies now appear live when the final stream payload is missed, without needing a page reload (#496) — @louis.

0.16.02026-06-19

Added2 entries

Personal sessions on mobile can now swipe from the opposite edge for session details, while filmstrip gestures respect swipe direction (#492, #495) — @jasper.
The personal agent sidebar now includes a tabbed session details inspector (#489) — @louis.

Changed4 entries

Feedback reports now submit optimistically and archive actions show a confirmation toast (#491) — @jasper.
Personal Brain and after-session memory guidance is stricter, so agents only save persistent files or memories when user intent is clear (#485) — @louis.
The personal sidebar home entry is now labeled New Session and uses the send icon (#487) — @louis.
Personal sidebar session lists now hide scrollbars while remaining scrollable (#494) — @jasper.

Fixed2 entries

Agent replies no longer reorder in the transcript when stream events arrive out of order (#493) — @jasper.
Mobile drawer swipes now work more reliably in Safari while preserving normal vertical page scrolling (#490) — @jasper.

0.15.02026-06-18

Added2 entries

Personal Brain now keeps restorable file versions, with restore tooling, caps, and runner guidance for safer durable writes (#483) — @jasper.
Mobile users can now swipe to open the main navigation drawer (#484) — @jasper.

Fixed4 entries

The onboarding wizard no longer resets when setting tool policies during the capabilities step (#481) — @jasper.
The three-dot menu in the brain file view no longer renders behind the file text (#480) — @jasper.
GitHub work repositories granted after connecting now resolve via a live lookup instead of being missed (#479) — @jasper.
Various mobile and phone UI fixes (#478) — @jasper.

0.14.02026-06-16

Added5 entries

OpenRouter Fusion models are now supported (#473) — @louis.
Messages sent mid-run now offer Steer, Queue, and Interrupt send-modes (#468) — @jasper.
Chat sessions can now be renamed via double-click or a right-click menu (#467) — @jasper.
Clarifying questions now always include a free-text "Other" answer (#466) — @jasper.
Screenshots dropped into chat are now attached to the Linear issue created from that conversation (#464) — @jasper.

Changed3 entries

Built-in skills now show distinct icons instead of a shared Sparkles placeholder (#475) — @jasper.
Memory lookup prompting tightened for more relevant recall (#474) — @louis.
Typed @mentions in the agent editor are no longer auto-converted as you type (#470) — @jasper.

Fixed4 entries

Near-identical person records in memory are now deduplicated (#477) — @jasper.
The thinking trace now separates distinct reasoning phases instead of running them together (#476) — @jasper.
The "Updated memory" card no longer overlaps the message footer (#469) — @jasper.
Fixed the personal billing redirect (#465) — @louis.

0.13.02026-06-12

Added8 entries

Agents can now create Linear issues directly from chat, preserving the relevant conversation context in the issue (#462) — @jasper.
Notion is now available as a first-party MCP integration with OAuth connection flow and tool discovery (#455) — @louis.
New users can now complete a Leo-led onboarding flow with updated personal onboarding screens and kickoff behavior (#451) — @louis.
Agents can now use the bundled YC knowledge skill (#456) — @louis.
Feedback reports can now include screenshots, which are uploaded and embedded inline in the linked Linear issue (#447) — @jasper.
The personal Memory surface is now URL-addressable, so selected memory paths can be opened and restored from direct links (#446) — @jasper.
Memory queries now support relative --since windows and query-less recent-memory listings (#444) — @louis.
Preview workspaces now show an early-preview badge in the sidebar (#457) — @louis.

Changed3 entries

Personal sidebar navigation was refreshed for the updated personal surface structure (#449) — @louis.
GitHub permission scopes were consolidated and renamed around clearer permission groups (#450) — @louis.
Signup welcome email delivery and unsubscribe handling were refactored for more reliable account email flows (#452) — @louis.

Fixed4 entries

Home-space image drag and drop now preserves sent images and shows them in the chat bubble after sending (#454) — @jasper.
Sessions now recover stale stream state after returning to a tab instead of staying stuck on outdated loading state (#448) — @louis.
Personal Brain writes are no longer silently dropped when the agent bundle exceeds the 80-file cap (#445) — @louis.
Chat scrolling now settles at the bottom when expected during active sessions (#431) — @jasper.

0.12.02026-06-11

Added10 entries

Brain files created or edited by agents now show as clickable attachments under assistant turns, and the company Brain can be opened directly at /brain/<path> deep links (#356) — @louis.
The personal Brain is now URL-addressable too, with nested file links restoring the selected file on reload and keeping the Brain navigation active (#426) — @jasper.
Personal Brain files created or edited during agent turns now link directly from the session transcript into the matching personal Brain path (#442) — @louis.
Personal sessions now support split-screen panes, so multiple chats can stay visible and be arranged side by side (#441) — @louis.
Agents can now start scoped subagents with selected tool grants, with subagent lifecycle events shown in the parent session transcript (#421) — @louis.
The personal Integrations tab now expands inline with connected account details, permission summaries, degraded-access warnings, disconnect actions, and GitHub repository refresh (#390) — @louis.
Personal recall can now be constrained by time windows, including query-less recent-session snippets, so agents can ask for memory from a specific lookback period (#416) — @louis.
Large pasted text is captured through the attachment pipeline, and oversized text attachments are materialized as sandbox file references instead of being inlined into every model call (#410) — @louis.
Session usage now updates live during streaming, long-running tool calls show an elapsed counter, and the context-window tooltip includes session cost (#395, #376, #399) — @louis.
Runner observability now records per-tool-call phase timings, Braintrust spans, PostHog rollups, and hidden sandbox hydration timing events for production latency diagnosis (#411, #417) — @louis.

Changed10 entries

Starting a session from the personal home prompt now switches to an optimistic session view immediately while creation and navigation finish in the background (#420) — @louis.
Runner sandbox hydration now overlaps bundle, skill, and setup work to reduce startup latency for sessions that need a sandbox (#434) — @louis.
Personal inbox cards now render markdown in their body and step text, matching the markdown contract used by the inbox tool (#438) — @louis.
Recall recap guidance is tighter, with cleaner formatting for recalled memories and follow-up summaries (#436) — @louis.
Preview deployment PR comments now include ?skipOnboarding, letting reviewers test fresh preview accounts without going through onboarding (#435) — @louis.
New personal agents now default to Leo on Kimi K2.6, with updated default capabilities, model selection, settings reset, and personal agent source normalization (#409, #425) — @louis.
Memory tooling now returns richer memory query and memory get output, clearer invalid-usage help, a tighter keeper kickoff prompt, and canonical guidance that durable facts belong in structured memory while agent/user.md is the always-loaded digest (#413, #418, #422, #428) — @louis.
Memory keeper passes now use Gemini 3.1 Flash Lite, surface what they stored, and appear in the transcript near the turn they followed instead of pinned to the bottom (#401, #397) — @louis.
Runtime details now stay collapsed by default for new sessions, including sessions with related parent or child metadata (#423) — @louis.
Slack support channel names now keep the per-workspace suffix at the end, for example <customer>-x-opencompany-<id8> (#394) — @jasper.

Fixed8 entries

Slash command session starts now route to the intended personal or company surface instead of opening the wrong session view (#432) — @louis.
Session lists now consistently sort by creation time across the company and personal sidebars (#437) — @louis.
Personal agent skill slash commands are available again in the composer slash menu and insert the right skill mention (#440) — @louis.
Personal-session inspector links now stay inside the personal surface instead of opening related sessions, session pages, or agent links under /company (#391) — @louis.
Personal runtime tool guidance now advertises the correct work/, personal-brain/, and agent/ roots and blocks generic file tools from editing structured memory paths (#414) — @louis.
OpenCode tool usage is now parsed from current step_finish events and charged back to the spawning session instead of silently appearing as zero-cost work (#402) — @jasper.
Runner shutdown and interruption handling now drains in-flight runs, persists abort/failure state more reliably, and avoids replaying interrupted deploy-time work (#378) — @louis.
E2B stream callbacks are guarded against detached rejections so stopping an opencode_coder command no longer crashes the runner process (#400) — @louis.

0.11.02026-06-10

Added6 entries

The new /personal experiment surface gives each user a private default agent with its own sessions, behavior editor, capability panels, context files, memory, tools, skills, integrations, and settings views (#353) — @louis.
Claude Fable 5 is now available end to end in the model catalog, with reasoning, image, PDF, long-context, and billing support (#384) — @louis.
Preview deployments can now complete Gmail and Google Calendar OAuth through a stable callback broker, making Google integration testing work across hosted preview URLs (#382) — @louis.
Public changelog entries can now include embedded images and short looping videos, with docs for recording, compressing, uploading, and referencing changelog media (#387) — @louis.
The account menu now shows when the currently running build was last deployed, with release details in the tooltip (#388) — @louis.
Personal agents now support plain @github for all installation repositories and repo-scoped @github/owner/repo mentions, with matching sandbox and coding-tool access (#392) — @louis.

Changed1 entry

Sandbox preparation no longer waits on optional developer-tool installation during the critical setup path (#393) — @louis.

Fixed1 entry

Personal recall searches no longer fail before running when the runner configures its statement timeout (#389) — @louis.

0.10.02026-06-08

Added3 entries

Paste, drag, drop, or pick screenshots, PDFs, and text/code files in the session composer; attachments are validated per model, stored in private Vercel Blob, rendered as compact cards, and inlined into runner model calls (#362) — @jasper.
Attached skills now appear in the composer slash menu as /<command> entries, using the skill's declared command frontmatter when present and inserting the matching @skill/<id> mention when selected (#350) — @louis.
Brain pages now show when the file was last updated directly in the header, and include a copy button for the page title plus current contents (#355, #358) — @louis, @jasper.

Changed2 entries

Session pages now use the session title as the browser tab title, making multiple open chats easier to tell apart (#354) — @louis.
Slack Connect support channels now use safer <customer>-<id8>-x-opencompany naming, include recovery sweeps for stuck provisioning, and have expanded setup and operations docs (#335) — @jasper.

Fixed3 entries

opencode_coder timeouts and command failures now preserve partial diffs, persist artifacts, return a resumable opencode session id, and report timeout status instead of dropping the work (#344) — @louis.
Long-running tool calls no longer trigger runner job replay and double billing after transient lease gaps; job lease TTL and lease-busy retry ceilings are now configurable (#345) — @louis.
Coding sandboxes now configure GitHub git credentials and best-effort ensure rg and bun are available, so plain git commands, repo search, and tests work more reliably inside sessions (#346) — @louis.

0.9.02026-06-05

Added7 entries

Gmail and Google Calendar are now first-party integrations — connect one or more Google accounts to give agents read-only Gmail and read/write Calendar tools, with per-account calendar selection (#319, #320) — @louis.
Switch the model for a single chat on the fly without changing the agent's saved default, in a redesigned composer with a toolbar tray for the agent and model selectors (#331) — @louis.
A context-window usage ring in the session top bar shows how full the model's context window is for the current turn (#332) — @louis.
Slack Connect — completing onboarding now provisions a private shared Slack channel with our team, surfaced as a "Connect on Slack" card on your workspace home (#278) — @jasper.
Tool calls that touch a Linear issue now show an external-link icon that opens the issue directly (#324) — @jasper.
Deferred tool discovery — capability tools and coding agents are no longer all preloaded; the model discovers them on demand via find_tools and runs them through a use_tool dispatcher, improving tool selection and shrinking the cached prompt (#321) — @louis.
New shared @opencompany/ui component package (shadcn / Tailwind v4) and a design-system app to document and showcase it (#323) — @louis.

Changed7 entries

Deferred-tool arguments now run through a validate → coerce → repair pipeline before dispatch, so smaller models' malformed tool calls are fixed automatically instead of failing (#328) — @louis.
Brain access scope is now enforced at the tool layer, so out-of-scope Brain writes are rejected up front with a clear error instead of being silently dropped at sync (#327) — @louis.
The new-session empty state no longer shows default prompt chips (#315) — @louis.
The /btw start toast no longer echoes the prompt; it shows just a confirmation and an Open action (#330) — @louis.
Faster navigation — removed the route-level session loader that flashed skeletons on fast nav (#303) — @louis.
Production deploys are faster and now recover gracefully from interrupted runner deploys (#310, #318) — @louis.
Runner model and tool calls are now traced via Braintrust's wrapAISDK defaults for richer, lower-maintenance observability (#304) — @louis.

Fixed3 entries

Unknown-tool and invalid tool-input calls are now recovered as tool results that steer the model back on track, instead of crashing the turn (#314, #325) — @louis.
Fixed a startup-status race that briefly showed "Stopped before finishing" on a freshly started session that was actually still running (#306, #312) — @louis.
Runner completions that contain only reasoning and no answer are now rejected instead of recorded as a successful turn (#300) — @louis.

Security1 entry

Hardened GitHub permission gating for agent shell commands, closing permission bypasses (#302) — @louis.

0.8.02026-06-04

Added8 entries

Real-time session sync and durable token streaming powered by Electric and Durable Streams, so chat updates propagate live and survive interrupted connections (#283) — @louis.
New opencode coding agent tool, letting agents run coding tasks through opencode (#266) — @louis.
Official PostHog MCP support with OAuth / Dynamic Client Registration, so agents can connect to PostHog without manual token wiring (#269) — @louis.
E2B sandbox compute is now metered and billed as agent usage (#264) — @louis.
Upload a custom profile picture for your account (#257) — @jasper.
Credit top-up analytics to track balance refills (#279) — @louis.
OpenCompany icon assets (#284) — @louis.
External skills support (V1) — attach a public GitHub repo or skills.sh page to an agent via an @skill mention; the skill is snapshotted workspace-side and materialized read-only into the runner sandbox at session start alongside built-in skills (#296) — @louis.

Changed9 entries

Session detail now paints instantly from local TanStack DB collections instead of waiting on the server (#290) — @louis.
Improved AI SDK prompt caching for lower latency and cost on repeated context (#275, #285) — @louis.
The agent self-edit skill now asks before making changes, stays lightweight, and better equips coding agents (#281) — @louis.
Onboarding call step reworked so its actions fit on a single screen (#263) — @jasper.
Web functions now run in the Frankfurt region for lower latency to European users (#286) — @louis.
MCP server tools are now lazy-loaded via per-server search_tools / use_tool meta-tools, so raw tool schemas are only sent to the model on demand instead of being injected on every turn — eliminating the per-turn token cost of large MCP catalogs like Slack (#298) — @louis.
Starting a new session from the root-route prompt now navigates instantly; session detail is assembled in-memory from the just-inserted rows instead of waiting on seven follow-up database reads (#295) — @louis.
E2B sandbox warm-up now begins at the start of a turn so the ~10 s cold-start overlaps the model's initial token stream rather than blocking the first tool call (#297) — @louis.
Runner parallel-session concurrency is now configurable via RUNNER_WORKER_CONCURRENCY (default 8); production limit raised to 40 (#294) — @louis.

Fixed10 entries

Coalesced durable-stream token appends to fix streaming lag in production (#288) — @louis.
Electric and Durable Streams now work in local development (#291), with producer linger and in-flight limits tuned for throughput (#292) — @louis.
Tool output previews are now redacted of secrets before being persisted to the database (#273) — @louis.
Approval stream state no longer gets stuck out of sync (#272) — @louis.
opencode now resolves the correct attached repository (#271) — @louis.
Linear save_* MCP tools are no longer misclassified as Admin-only (#267) — @louis.
@mention pills no longer break after an agent edits itself (#268) — @louis.
A run is never replayed after its lease expires, and onboarding dedup is hardened against duplicate runs (#262) — @louis.
The Render deploy trigger is now resilient to empty API responses (#287) — @louis.
Session status and error no longer flicker blank at the start of a session; the live stream only takes authority over the Postgres snapshot once it has received a status-bearing event (#299) — @louis.

0.7.02026-06-03

Added5 entries

Agents can now pause mid-run to ask structured questions with single-select, multi-select, and "Other" answers, then continue from the user's response (#255) — @louis.
Onboarding agents now use structured questions and receive signup names in their starting context, making first-run setup more personal and less brittle (#259, #260) — @louis.
Apify-backed Instagram and TikTok tools for public profiles, posts, reels, videos, comments, search, and async social scraping jobs (#253) — @louis.
Searchable, provider-grouped model picker with Capability, Speed, and Cost ratings so models are easier to compare at a glance (#254) — @louis.
Day-one Brain wiki and private agent agent/soul.md scaffolding, giving new workspaces a cleaner shared knowledge structure and better self-configuration defaults (#251) — @louis.

Changed2 entries

Onboarding call booking now uses one booking-aware button, the official Cal.com embed bootstrap, and clearer new-tab labeling (#249, #250, #258) — @jasper.
Public changelog entries now include author attributions (#256) — @louis.

Fixed4 entries

Chat scroll-to-top behavior after sending a message is now stable during streaming and viewport resizing, with less jitter and less fighting user scrolls (#248) — @jasper.
Paused sessions now refresh stream tokens correctly, and settings hydrate consistently after reloads (#252) — @louis.
Editing a markdown title or header no longer flickers while typing (#201) — @jasper.
Naming a new Brain file from its tab now updates the sidebar tree immediately (#198) — @jasper.

0.6.02026-06-02

Added11 entries

Guided first-run setup — after signing up you land in a live setup conversation that helps configure your workspace, instead of an empty editor (#241) — @louis.
New /btw command to start a new session without leaving your current view (#243) — @louis.
New /clear command in the chat composer (#231) — @louis.
"Run now" button to trigger a scheduled agent immediately (#236) — @louis.
Agents can now pull content from TikTok and Instagram (#239) and YouTube (#233) — @louis.
Pin sessions to the top of the sidebar (#213) — @jasper.
Live product status indicator in the sidebar, linking to the public status page (#224) — @louis.
Agents can now adjust their own schedules (#217) — @louis.
Workspace tool permissions — allow, ask, or deny which tools and integrations an agent can use, with approval prompts that are remembered between runs (#229) — @louis.
Live reasoning support for Kimi models (#226) — @louis.
Provider logos for xAI and MiniMax in the model picker (#232) — @jasper.

Changed5 entries

Agent runs no longer break when an integration like Linear or Slack is turned on but not yet connected — the agent now asks you to connect it instead (#238) — @louis.
When you send a message, it scrolls to the top of the chat, matching the ChatGPT/Claude experience (#219) — @jasper.
The copy button now stays visible on assistant replies and appears on hover for your own messages (#222) — @jasper.
The onboarding call step now has a clear "Skip for now" button (#220) — @jasper.
The "reconnecting" banner no longer appears in chat during normal streaming; connection status now lives in the inspector instead (#242) — @louis.

Fixed6 entries

Linear no longer shows as disconnected after you reconnect with valid credentials (#234) — @louis.
Agent turns that quietly stop mid-task are now flagged instead of being recorded as a silent success — important for scheduled and unattended runs (#225) — @louis.
Sessions now recover when a response stream gets interrupted (#223) — @louis.
Rapidly toggling a session's pin no longer causes glitches (#218) — @jasper.
Restored the after-session hook menu when typing at the start of a line (#227) — @jasper.
The hook suggestion popup now stays beneath the model dropdown (#230) — @jasper.

0.5.02026-06-01

Added7 entries

Cron-based agent schedules — agents can now be scheduled to run on a recurring basis (#208) — @louis.
Read-only X (Twitter) hosted tool for agents (#212) — @louis.
Agent self-update skill, allowing agents to evolve their own .agent definition (#210) — @louis.
Dark mode support (#186) — @louis.
Grok model support via the AI gateway (#206) — @louis.
MiniMax and Kimi gateway models (#193) — @louis.
Client-side felt time-to-first-token measurement (#192) — @louis.

Changed5 entries

GitHub repositories are now decoupled from the amp tool, making repository bindings more flexible (#204) — @louis.
Lazy GitHub sandbox initialization is more solid and reliable (#207) — @louis.
Vercel Skew Protection is enabled via custom deploymentId for safer deploys (#205) — @louis.
Onboarding call booking step is now skippable (#191) — @jasper.
Assistant turn duration is now shown next to the copy button (#187) — @jasper.

Fixed6 entries

Agents no longer disappear from the list after creating or editing a new agent (#203) — @jasper.
Renaming an agent no longer resets it to "Untitled agent" (#194) — @jasper.
Vercel deployment ID length is correctly handled (#211) — @louis.
Session archiving is now optimistic and instant instead of waiting on the server (#200) — @jasper.
Feedback dialog auto-closes after a successful submit (#190) — @jasper.
Page no longer rubber-bands at the fold due to overscroll-behavior fix (#189) — @jasper.

0.4.12026-05-30

Changed3 entries

Runner job worker now wakes immediately on enqueue instead of waiting for the next poll interval, reducing time-to-first-token for agent sessions — @louis.
Web dispatch no longer blocks on analytics flush; PostHog capture and runner dispatch now run concurrently — @louis.
Workspace context loading collapses user, workspace, and role into a single joined query, and session submission runs the balance check and auth lookup concurrently, cutting pre-dispatch database round-trips — @louis.

Fixed1 entry

Runner event serialization no longer crashes when created_at arrives as a string from the raw lease-write path; the value is now coerced to a Date at the source — @louis.

0.4.02026-05-28

Added7 entries

Workspace integration resource bindings with encrypted credential storage, multi-connection GitHub repository binding, and runtime validation for bound resources.
Linear and Slack MCP support, including workspace setup flows, agent editor tool gating, runner-side MCP execution, and .agent file configuration.
Delegated agent sessions through @agent/<slug> mentions and a delegate_to_agent tool, with child-session tracking and parent usage rollups.
Repo-scoped GitHub auth for agent shell commands and AMP runs, with ephemeral credentials and output redaction.
GPT 5.2 Codex support and configurable AMP execution modes.
Agent editor support for markdown headings, ordered and unordered lists, and automatic conversion of typed or pasted @/# mentions into pills.
A richer chat composer with quick-start chips, auto-resize, stop controls, copy-message actions, drag-and-drop and paste handling shells, and improved keyboard hints.

Changed3 entries

Local setup now keeps WorkOS authentication redirects on localhost:3000 while continuing to use ngrok for GitHub and other public integration callbacks.
Production web deploys now rely on the GitHub Actions release workflow instead of Vercel Git-triggered deployments.
AMP command handling now avoids rejected stream-JSON mode combinations and falls back to sanitized plain output when structured results are unavailable.

Fixed5 entries

GitHub integration sync in production now avoids unsupported Neon HTTP transactions and normalizes legacy agent configs.
Sandbox repository clones, AMP publishing, and authenticated shell commands now handle GitHub App credentials more reliably.
Session chat no longer allows duplicate submissions while an assistant is responding and no longer leaves failed sessions stuck in a thinking state.
The New agent button now shows immediate pending feedback, blocks double-click creation, and surfaces creation failures.
Linear MCP credential setup now reports missing encryption-key configuration with clearer diagnostics.

0.3.02026-05-27

Added6 entries

Deterministic edit_file support for hosted agents, giving runner sessions a safer and more reviewable way to modify files.
Expanded hosted Exa tools and clearer search schema guidance for agent research workflows.
Common Vercel AI Gateway model presets across the editor, runtime configuration, and billing calculations.
Model and turn-cost analytics for better visibility into session usage and credit spend.
Signup welcome emails, session-aware feedback reports, and an onboarding call booking step.
A pulsing sidebar indicator for active sessions.

Changed4 entries

Runner internals now separate session lifecycle, job leasing, tool dispatch, model streaming, and usage recording for more reliable hosted runs.
Agent file parsing, mention handling, runtime types, and after-session behavior now live in the shared agent runtime package used by both web and runner.
GitHub sync jobs now have retry and sweeper support to recover pending workspace file updates more reliably.
Development and review tooling now includes CodeRabbit configuration and more consistent local script environment loading.

Fixed4 entries

Blank user messages now survive session payload handling and runtime event rendering.
Agent detail mention caches no longer leak across agent boundaries.
Hosted tool fan-out is capped to avoid runaway parallel tool execution.
GitHub integration callback diagnostics now expose enough detail to troubleshoot failed setup flows.

0.2.02026-05-26

Added7 entries

Git-backed agent editing, including file-backed .agent sync, rename propagation, and visible GitHub sync state.
Hosted agent sessions with root-route session start, run controls, archive lifecycle, automatic sandbox pausing, and first-message session titles.
Agent tools for hosted Exa search and lightweight web fetching.
WorkOS organization workspace support, signup onboarding, default workspace credits, and personal environment setup.
Brain context files for agents, plus a more complete workspace file experience.
Session usage, billing, analytics, and cached-token tracking so teams can understand credit spend.
Public docs, public changelog, and in-app feedback intake.

Changed5 entries

Workspace sessions now load from cache first and route transitions respond faster.
Session UI now better explains reasoning model choice, token usage, tool replay, and follow-up state.
Workspace loading and agent navigation states are more predictable.
Release automation now coalesces CI, syncs Inngest production functions, records GitHub deployments, and runs on Blacksmith.
Observability now covers agent sync, launch events, runner logs, and Better Stack error context.

Fixed5 entries

AuthKit sign-in and sign-up redirects now use full document navigation when required.
Runner sandboxes recover more reliably from E2B lifecycle and tool failures.
Brain new file and new folder actions no longer throw on click.
Direct session title generation and tool replay behave consistently.
Production release smoke checks and Render release waiting are more reliable.

Removed1 entry

Persisted tool delta events and the sidebar running badge to reduce noisy state.

0.1.02026-05-20

Added3 entries

Initial Next.js app prototype with sidebar navigation, Agents, Brain, Inbox, Settings, and session surfaces.
Early WorkOS authentication, branchable Neon database setup, and persistent agents.
Turborepo workspace structure with Bun, shared database package, CI, formatting, tests, and secret scanning.