Skip to content
Critique
Log inInstall

Early access · machine-local verifier

Check the change before it ships.

Critique reviews the code on your machine before you commit it. It explains what it checked, what failed, and what to do next.

Install the CLI Open hosted Code Read the protocol

runs on your working treeCritique Code harnessstructured handoff

finish runfin_7f3a

$ critique finish --intent "Stop duplicate charges" --repair pack --json

01change reconstructedready
02reproduction + limitsattached
03repair packageprepared
caller decides next stepexit 2
Where it runs
Your working treeStaged, unstaged, deleted, and untracked changes.
Verifier
Fresh model familyIndependent session, reset workspace, captured evidence.
Agents
Claude Code · CodexCursor, Copilot, and any agent can call the same CLI.
Output
finish.v1 JSONOutcome, evidence, limits, and exit codes.

Supported agents

Works with the agents you already use.

  • CursorCursor
  • Claude CodeClaude Code
  • CodexCodex
  • opencode
  • AntigravityAntigravity
  • Hermes AgentHermes Agent
  • OpenClawOpenClaw
  • LobeHubLobeHub
  • ClaudeClaude
  • OpenAIOpenAI
  • Kimi
    Kimi
  • ManusManus
  • CherryStudioCherryStudio
  • AmpAmp
  • GithubCopilotGithubCopilot
  • CopilotCopilot
  • Gemini CLIGemini CLI
  • GeminiGemini
  • Kilo CodeKilo Code
  • JunieJunie
  • ReplitReplit
  • WindsurfWindsurf
  • ClineCline
  • OpenHandsOpenHands
  • RooCodeRooCode
  • GooseGoose
  • TRAETRAE
  • ZencoderZencoder
  • Coze
    Coze
  • Dify
  • QoderQoder
  • QwenQwen
  • AzureAzure
  • OpenWebUI
    OpenWebUI
  • CodeFlickerCodeFlicker
  • FlowithFlowith
  • 01.AI
  • MetaAIMetaAI
  • n8nn8n
  • NotionNotion
  • CodeBuddyCodeBuddy
  • DevinDevin
  • CursorCursor
  • Claude CodeClaude Code
  • CodexCodex
  • opencode
  • AntigravityAntigravity
  • Hermes AgentHermes Agent
  • OpenClawOpenClaw
  • LobeHubLobeHub
  • ClaudeClaude
  • OpenAIOpenAI
  • Kimi
    Kimi
  • ManusManus
  • CherryStudioCherryStudio
  • AmpAmp
  • GithubCopilotGithubCopilot
  • CopilotCopilot
  • Gemini CLIGemini CLI
  • GeminiGemini
  • Kilo CodeKilo Code
  • JunieJunie
  • ReplitReplit
  • WindsurfWindsurf
  • ClineCline
  • OpenHandsOpenHands
  • RooCodeRooCode
  • GooseGoose
  • TRAETRAE
  • ZencoderZencoder
  • Coze
    Coze
  • Dify
  • QoderQoder
  • QwenQwen
  • AzureAzure
  • OpenWebUI
    OpenWebUI
  • CodeFlickerCodeFlicker
  • FlowithFlowith
  • 01.AI
  • MetaAIMetaAI
  • n8nn8n
  • NotionNotion
  • CodeBuddyCodeBuddy
  • DevinDevin

System facts

Small by design.
Explicit by default.

The finish pass is deliberately compact: a few named stages, a visible execution path, and one result the caller can act on. Every tick below is a real part of that contract.

01 / THE FINISH CONTRACTFour facts. No mystery math.one tick = one declared unit · the rail stays honest
Countable system facts in the Critique finish contractFour horizontal tick rows show three protocol stages, two execution paths, four finish fields, and one caller handoff.PROTOCOL STAGES3EXECUTION PATHS2FINISH FIELDS4CALLER HANDOFF1ONE TICK = ONE DECLARED UNIT · DOT MARKS THE RAIL END
01Protocol stages
3Inspect · decide · continue
02Execution paths
2Local or clean cloud machine
03Finish fields
4Outcome · evidence · limits · exit
04Caller handoff
1A structured result, not a dashboard ritual

Your agent writes the change.
Critique checks it.

/ 01 — Position

It looks for real problems, runs useful checks, and returns a result your agent or your teammate can act on. One separate opinion, in the loop where the work happens.

The protocol

One call.
A second set of eyes.

/ 02 — Protocol

Critique returns a judgment, the evidence behind it, and a stable handoff your agent can consume without scraping a dashboard.

01

Inspect

Reconstructs the change in a separate workspace and tests the claims behind it instead of re-reading the diff.

critique finish --intent "…"
02

Decide

Returns one clear outcome, the evidence that produced it, and the limits of that evidence.

outcome · evidence · limits
03

Continue

Hands a structured result back to whoever owns the next move — your agent, or you.

exit 0 · exit 2

The handoff

Your agent keeps the wheel.

Critique starts with the task, the exact diff, and the files your agent touched. Local runs follow the change into nearby code and tests in a separate workspace. Cloud runs upload a protected snapshot of the full working tree — uncommitted changes included — then run it and return the test evidence.

/ 03 — Handoff
Critique Code keeps the implementation loop and its evidence in view.

Built for the caller

Built for the agent making the call.

No dashboard ritual. Your coding agent runs one command, reads structured output, and decides what happens next.

See the CLI contract
agent callfresh context
$ critique finish --intent "Stop duplicate charges" --repair pack --json

→ outcome repair_ready
→ evidence reproduction + limits attached
→ repair patch package available

CALLER DECIDES NEXT STEP
structured handoffexit 2

Three commands

Three commands.
One independent check.

/ 04 — Commands

Install it once, add it to the agents you already use, then call Critique whenever a change needs an independent decision.

01

Authorize once

critique login

Scoped machine credential, one device approval.

Human step
02

Connect agents

critique integrate --agent all

Claude Code, Codex, Cursor, Copilot, and other agents.

One command
03

Check a change

critique finish --intent "…" --repair pack --json

Structured result, plus a repair package when one is warranted.

Agent loop
CRITIQUE CODE
C
HARNESS OUT

The harness

Critique Code is an agentic coding harness.
Write the change. Prove it. Same loop.

Same class as Codex and Claude Code. Any model: your providers or Critique Inference. Run it from the agents you already use, or drive it yourself with critique-code. It happens to be lethal at review because it actually writes and tests the code.

See the harness

Install the CLI

Put an independent check in every agent loop.

npm install --global @critiquedotsh/cli

v0.1.1Node 20+macOS · Linux · Windows

Then authorize one device and add Critique to the agents you already use.

critique login critique integrate --agent all
Download package ↓View on npm ↗

Keep exploring

The CLI is the product.
These are its essentials.

00Hosted Code

Open a capability-gated CritiqueCode session in the browser. Pick up the work from any device.

01Critique Code

The agentic coding harness. Controller-owned loop. Package @critiquedotsh/harness, binary critique-code.

02CLI contract

Installation, local-change review, skills, model selection, repair modes, and finish.v1.

03Pricing and models

Compare plans, bring your own key, and choose the verifier you trust.

04Blog

Research, comparisons, and the case for independent completion checks.

05Inference API

Use Critique’s OpenAI-compatible model API wherever your product needs it.