About Critique
Built for code that has to hold.
AI agents write code. Critique verifies it in a sandbox before you merge. Critique is an Automated code verification platform for GitHub pull requests — not another comment bot on your diff. We gate untrusted changes, run evidence-backed review, enforce merge policy, and repair with proof on one Change Passport per pull request.

Change control
Not just a code review CI tool.
v4 treats the Change Passport as the product object: provenance, risk, evidence, policy decisions, remedy proof, and memory on one merge-boundary record. Review runs are commit-level evidence generators inside that passport.
Linked context
Chat, review, and remedy share repository state instead of re-explaining the PR.
Checkpoint first
Policy gates run before the review swarm spends on untrusted shape.
Visible artifacts
Findings, scope, and handoff contracts stay on the trail.
Bounded execution
Automation stops at agreed scope — not an open-ended rewrite.
Gate
- When
- Before the review queue
- Check
- Critique / Checkpoint
- What it does
- Fast, cheap firewall rules
Review
- When
- During the run
- Check
- Critique / Review
- What it does
- Multi-model specialists + lead verdict
Merge
- When
- After evidence
- Check
- Critique / Merge Policy
- What it does
- Enforceable merge rules with proof
The loop
Chat, review, then repair.
We started with review because finding the truth has to come before asking automation to act. Checkpoint can gate the PR before the review swarm spends.
Chat
Ask repository questions without starting from zero.
Review
Run a grounded PR critique with specialist lanes.
Remedy
Turn eligible findings into validated branch updates.
- dry-runObserve policy impact without blocking contributors.
- warnSurface risk while still allowing the review to continue.
- blockStop untrusted or malformed PRs before full automation starts.
Principles
How we decide what ships.
Evidence over vibes
traceA finding should carry enough context for an engineer to verify it fast.
Trust is earned
gateAutomation must show its work, obey scope, and accept when it should stop.
Code before theatre
quietThe interface stays quiet because the work is already complicated.
Taste under pressure
craftDeveloper tools should feel precise even when the system is busy.
Local control
policyTeams choose policies, models, and where execution runs.
Autonomy with boundaries. That is how trust is earned.
Shipping now
Live in beta, shipping weekly.
- Control Board: gate, policy, memory, delivery
- Change Passports per pull request
- GitHub App review runs with evidence contracts
- Workspace: chat, review, builder
- Checkpoint / Agent Firewall on the merge boundary
- Remedy with proof bundles
Ship log: changelog · Skill Marketplace
Start here
Better code. Shipped with confidence.
Connect GitHub, run your next PR through review, and use Remedy when the fix has enough evidence to act.
Contact
Product and account help: support@critique.sh. Founder: ray@critique.sh.