A critic in the session
CLI and MCP into Claude Code, Codex, Cursor, Grok Build, and OpenCode — Cerebe reviews what the agent writes, against your repo's rules, so slop is caught before the pull request exists.
Cerebe is the quality engine for AI software development — a critic beside the agent while it writes, and an independent panel on every pull request. One verdict, bound to the commit. Start advisory; require the check when you want it to gate merge.
CLI and MCP into Claude Code, Codex, Cursor, Grok Build, and OpenCode — Cerebe reviews what the agent writes, against your repo's rules, so slop is caught before the pull request exists.
The hosted GitHub App runs independent critics from rival model families and posts one verdict as a check.
Advisory posts a verdict on every pull request. Required means that check must be green to merge. Same product, two postures.
Install
The CLI is free to run on any repository. The hosted GitHub App — that's the platform panel. Require the Cerebe check when you want enforcement.
curl -fsSL https://raw.githubusercontent.com/momentiq-ai/cerebe/main/install.sh | sh cerebe --version cerebe install
Install is the workstation. The GitHub App is the platform.
How the gate works
Build, test, schema, and format run before any model. If a mechanical check can decide, no LLM is spent — and a compiler won't be argued out of a type error.
Independent critics from rival model families review the residual risk in parallel, each on its own — different lineages catch different problems, and the verdict is theirs together.
Opinions in, ship-or-block out. No network, no model in the decision — the same kernel on your laptop and in the hosted check.
The verdict and every finding are bound to the commit SHA and diff hash. Amend the commit and the gate re-runs. Bypass is loud and append-only.
Declared objectives are derived from the source that declared them, hashed to it, and bound to a real route or test — intent that survives the session.
The hosted GitHub App posts the merge check itself — no workflow line to delete. Local review is a convenience; the platform is the authority.
Every part of the engine
Independent critics from rival model families review each change in parallel under a min-complete-quorum policy. Severity tags shape the verdict, and if a critic times out the gate continues with the ones that completed and records what was evaluated.
Every verdict is bound to the commit SHA and diff hash — the proof travels with the commit, and if the diff changes the gate re-runs. Append-only, so the history stays a forensic trail.
Independent findings become one deterministic decision: severity tiers set what holds the merge and what advises, applied the same way on the workstation and in the hosted check.
A blocking finding holds the merge when a second independent critic corroborates it at the same place — decisive and steady. Findings that stand alone are kept on the record as dissents.
When a change touches the gate's own policy, the critic evaluates it against the trusted base — every change is judged by rules it didn't set, so the gate stays honest.
Every decision and every override is an audited event with a recorded reason, kept append-only — a complete trail you can hand a reviewer.
Whatever model families you already run, Cerebe installs the same way and works across them. Your stack, your call.
An agent's context anchors to a work item and picks up again later — the long arc keeps its reasoning across sessions, machines, and handoffs.
The consequential decisions behind a change are recorded, classified, and tracked to resolution, so the reasoning lives on the record.
A critic on the workstation. A panel on the platform. Install for advisory. Require the check when you want it to gate merge.