Evaluation: crit - Human-in-the-Loop Review Interface for AI Agents
Resource Type: Open Source CLI Tool Author: Tomasz Tomczyk (@tomasz-tomczyk) Date: 2026-06-09 (v0.16.2) URL: https://github.com/tomasz-tomczyk/critEvaluation Date: 2026-06-10 Evaluator: Claude Sonnet 4.6
1. Content Summary
crit is a local review interface designed to sit between a developer and an AI agent. It provides an adaptive UI for reviewing AI-generated output across four modes:
- Markdown/plan files: Rendered with inline line-or-range comments before the agent executes
- Git diffs: Split/unified view with round-to-round comparison (delta between agent iterations, not just commits)
- Live web apps: Proxies a running local server and injects a DOM-anchoring overlay for spatial comments mapped back to source
- Static HTML: Direct render with the same overlay system
The loop: agent produces output → developer opens crit → leaves inline comments → agent reads those comments → crit shows what changed round-to-round.
Key features:
- Live threads (continuous back-and-forth conversation attached to a specific line, not one-shot)
crit comment src/file.go:42 'message': programmatic CLI for agents to inject their own notesPOST /api/agent/request: HTTP endpoint for agents to push review requests to the human- GitHub PR bidirectional sync (read/write PR comments via
ghCLI) crit install claude-code: first-class Claude Code integration (writes project config snippets)- Async shareable review URLs for remote collaboration
Tech stack: Go 1.26+ backend, JS/TS frontend embedded in binary. No database, no cloud, no telemetry.
2. Initial Scoring: 4/5 (High Value)
| Score | Meaning | Action |
|---|---|---|
| 5 | Critical: integrate immediately | < 24h |
| 4 | High Value: major improvement | < 1 week |
| 3 | Moderate: useful addition | When time available |
| 2 | Marginal: minimal mention or skip | - |
| 1 | Low: reject | - |
Justification
Strengths:
- Direct Claude Code integration (
crit install claude-codeis a documented first-class command) - Fills a genuine gap: no existing tool shows round-to-round diffs within an AI session
- DOM-anchored spatial comments for live web apps, with no equivalent in standard PR tooling
- Active maintenance: 477 stars, v0.16.2 shipped June 9 2026, 665 total commits
- MIT license, no telemetry, no cloud dependency
- The CLAUDE.md in the repo is a symlink to AGENTS.md, meaning the maintainer understands Claude Code conventions
- Addresses a friction pattern that shows up in Claude Code sessions: "how do I give structured feedback on a plan before Claude executes it?"
Why 4/5 and not 5/5:
- Single maintainer (bus factor), contributor depth unclear
livemode is explicitly marked "feedback wanted", not yet stable- Go 1.26+ build requirement (may cause issues building from source)
- No visible test suite in README
- Documentation functional but thin (no dedicated docs site)
3. Technical Analysis
Round-to-Round Diffing: The Unique Value
Standard diff tools (GitHub PRs, git diff, IDE diffs) show delta between commits or branches. crit shows delta between agent iterations within the same review session. That specific capability does not exist in any other tooling short of manual bookmarking.
For Claude Code users specifically: when you run Claude multiple times on the same task and want to understand "what did it change from attempt 2 to attempt 3?", crit gives a structured answer.
Plan Review Before Execution
The markdown mode enables reviewing plan.md or similar files inline before the agent proceeds. Without crit, this requires manually opening the file and passing feedback through the prompt. With crit, comments are attached to specific lines and readable by the agent via the CLI comment API.
GitHub PR Sync
crit can read and write GitHub PR review comments bidirectionally via gh CLI. This bridges the loop: Claude Code opens a PR, reviewer leaves inline comments in crit, those comments sync to the PR, Claude Code reads them via the PR context.
4. Maturity Signals
| Signal | Value |
|---|---|
| Stars | 477 (now 834 as of 2026-07-28) |
| Forks | 36 |
| Open issues | 7 (6 feature requests, 1 bug) |
| Latest release | v0.16.2, 2026-06-09 |
| Total commits | 665 |
| License | MIT |
| Topics | ai-agents, llm, agentic-coding, code-review, cli, developer-tools |
The issue queue is healthy: one recent bug (@ filename handling, minor), the rest are scoped feature requests with "help wanted" labels. Release cadence is active.
5. Installation
# Homebrew (macOS/Linux)
brew install crit
# Go
go install github.com/tomasz-tomczyk/crit@latest
# Binary, Nix, Docker also availableBasic usage:
crit # Auto-detects uncommitted changes in repo
crit plan.md # Review a specific file before agent executes
crit http://localhost:3000 # Proxy and review a running web app
crit landing.html # Review static HTML with spatial comments
# Claude Code integration setup
crit install claude-code # Writes config snippets into the project6. Relevance to Claude Code Users
Direct and explicit. The tool is built for this workflow:
crit install claude-codeconfigures it for Claude Code sessions out of the box- The HTTP API (
/api/agent/request) is designed for Claude Code to be the receiving agent - Round-to-round diffs map directly to how Claude Code iterates on the same task across multiple runs
- PR sync bridges the gap between Claude Code-opened PRs and structured human review
For guide readers: crit addresses a friction point that comes up in multi-iteration Claude Code sessions. The current guide covers plan review and feedback patterns, but has no tooling recommendation for structured inline review with round-to-round comparison. This fills that gap.
7. Red Flags
No serious issues. Minor notes:
- Single maintainer: Active and engaged, but bus factor is real at 477 stars
- Go 1.26+ to build from source: Homebrew/binary installs are unaffected; build-from-source users may hit this
- Live mode not stable: Issue #557 explicitly labels it as seeking community feedback
- One open bug:
@in filenames (#656, 2 weeks old, unresolved but scoped) - No test suite visible in README: No CI badge or test command documented
agent_cmdis global-only: Deliberate security decision (prevents malicious repos hijacking your agent), but limits per-project agent configuration
No security concerns. No telemetry. No supply chain issues visible.
8. Integration Decision
INTEGRATE at score 4/5: mention in the guide under human-in-the-loop and review tooling.
Where to integrate:
- Section covering agentic workflows and human oversight patterns (the guide section on iteration and review)
- Third-party tools ecosystem section
- Potentially: a callout in the multi-agent / plan review workflow sections
Format: Short mention with link and what it uniquely solves (round-to-round diffs, plan review before execution, DOM-anchored web review). No need to document the full feature set, just link to the GitHub repo. One-liner pitch: "structured inline review interface between you and Claude Code, with round-to-round diffs across iterations."
Not a priority over core guide work, but a clean addition during the next third-party tools pass.
9. Final Metadata
Score: 4/5 Decision: Integrate Confidence: High
Next action: Add mention in third-party tools / agentic workflow sections on next guide update pass.
Archive: docs/resource-evaluations/crit-human-in-loop-review.md