Your coding agents keep making the same mistakes. Company Brain reads what Claude Code, Codex and Cursor already did, writes the rules and lessons, and every agent on the team is told to read them before its next session.
An illustrated example. Each entry the brain writes names where it came from.
npm run build:staging reads the database at build time. Two attempts hit the same error.
Your agents keep working where they already work. Each session is told to read what the team learned before it starts.
One MCP endpoint for every agent on the team. Each session is told to read the brain before it starts.
claude mcp add --transport http companybrain https://companybrain-board.vercel.app/mcpStart a new session. Claude Code opens a GitHub sign-in the first time it connects.
Sign in with GitHub, open Sources and index a repository. Every read uses your own GitHub authorization.
Past Claude Code, Codex and Cursor conversations, from your machine. It strips common secret formats before anything leaves, shows you the list first, and sends nothing until you add --send.
curl -fsSL https://companybrain-board.vercel.app/import.mjs -o companybrain-import.mjsnode companybrain-import.mjsNeeds Node 22 or newer. When the list looks right, create an import token in the app, run it again with --send and paste the token when asked. Then have the Librarian learn from them:
Auto-build runs the Librarian over your indexed repos and sources. It writes the rules, lessons, processes and skills that are missing and never replaces what a person wrote. Turn on Build every day to keep it current.
Four layers, one brain. What your agents do is captured, kept as rules and lessons with their sources, improved by the runs that use it, and served to every agent you run.
Connect your GitHub repos and import past Claude Code, Codex and Cursor sessions, with common secret formats stripped on your machine first. Slack, Linear and Notion are optional. Every new agent run adds what it learned, so a fix stops disappearing when the chat closes.
The Librarian writes your team's rules, lessons, processes and skills as typed records. Each one links to the sessions, files and docs it was built from and keeps its recent versions, so you can see what changed and why.
The Reflector reads how runs went, turns repeated failures into lessons and proposes better skills. A change is applied only if it beats the current version when replayed against past goals, or you approve it. Until enough runs are graded to test against, every change waits for your approval.
One MCP endpoint for Claude Code, Codex, Cursor and more. Each session is told to read the brain before it starts. Connect up to 20 MCP servers from a directory of 135, behind the same endpoint, with credentials kept on the server.
A timeline of every entry the brain wrote, and where each one came from.
Tasks it suggests to an agent or a person, tracked until the work is accepted.
Questions an agent should not answer alone, escalated to a person. The agent stops and waits for the ruling.
Connect up to 20 MCP servers from a directory of 135, behind the same endpoint, with credentials kept on the server.
Meetings, Slack threads, Linear projects and the decisions behind them. Ask in plain words, get the answer with its sources.
CRM notes and account history in the same brain, so the next call starts from what the last one learned.
The importer strips common secret formats before anything leaves your computer, shows you the list first, and sends nothing until you say so. The server strips them again.
It reads a connected server only when you turn it on for that server, and only calls tools marked or named as read-only. Every gateway call is logged.
Entries the brain writes cite the sessions, files and docs they were built from.
A change to an existing skill, process, rule or role is applied only if it beats the current version on past goals, or you approve it. Until enough runs are graded to test against, every change waits for your approval. New entries are added alongside and never replace what a person wrote. Any applied change rolls back in one click, and each entry keeps up to 20 versions from the last 90 days.