The system of record for product work — built for teams whose builders are agents.
Requirements, epics, acceptance criteria, and feedback live in one ledger. People decide; agents build; every action is attributed to both the person and the token that acted.
What it is
Pipebrain is a work ledger, not an orchestrator. It records intent, holds claimable work, and receives reports. Agents connect through an MCP server and consume the same API as every other client, so no surface quietly gains a capability the others lack.
- Requirements are living documents. Versioned, approved by a person, and delivered to every agent session automatically — always at the approved revision.
- Criteria are the contract. Each slice of work carries checkable acceptance criteria; machine-verifiable ones carry evidence, human ones require a signed-in person. The server enforces the gate.
- Attribution is structural. Every write records the acting person and the token used. Human-only actions refuse tokens outright.
- Secrets stay out of transcripts. Secret values move only through the CLI and the web UI — no agent tool can read one back, by design.
Install
The agent server and the CLI are public npm packages:
npx -y pipebrain-mcp # the MCP server (stdio) — register it with your agent
npm i -g pipebrain-cli # the ppb command — doctor, secrets, assets
Register the server with Claude Code in one command. The server has no default
address and no default organization, so three values are passed here — set
PIPEBRAIN_API_URL to https://api.pipebrain.dev, export your
PIPEBRAIN_TOKEN first, and replace your-org-slug with your
own, since a call that names no organization is refused:
claude mcp add pipebrain -s user -e PIPEBRAIN_API_URL=https://api.pipebrain.dev -e 'PIPEBRAIN_TOKEN=${PIPEBRAIN_TOKEN}' -e PIPEBRAIN_ORG=your-org-slug -- npx -y pipebrain-mcp@latest
Then ppb doctor to check the wiring; with no organization named in your
own shell yet, it also prints the ones you may use. Full setup — Claude Desktop, and
the remote endpoint for clients that cannot run a local server — lives in the in-app
guide after sign-in.
Documentation
The three documents Pipebrain is built from, run by, and instructs its agents with are published here in full:
- Requirements — the V1 specification: the domain model, the MCP tool catalog, the API, and the data and security invariants.
- Delivery workflow — how work moves from intent to done, and where a person has to decide.
- Agent doctrine — the rules every agent session is given, identical in every organization: how to write for the person deciding, what an agent decides alone, and how each kind of record is written.
Access
Pipebrain is invitation-only. There is no open signup: a member of an organization invites you by email, and the invitation is what creates your account. If you were invited, follow the link in your email; if you'd like an invitation, contact the operators.