Crit: Local-first, single-binary CLI for reviewing agent plans and diffs
Crit is a local-first review tool packaged as a single binary. It opens your file or code diff in a browser with a GitHub-inspired interface, lets you leave comments, then your agent acts on the feedback automatically. Repeat until satisfied — all before hitting GitHub.
Core loop: comment, agent updates, diff refreshes
Drag across a range of lines, leave comments, hit Finish Review. The agent picks up the feedback, edits the file, and Crit reloads with the updated diff. Previous comments stay visible; resolve only when you're satisfied. Works for both plans and code — same UI, beautifully rendered markdown with syntax highlighting for 190+ languages.
Install in one command
$ brew install crit $ go install github.com/tomasz-tomczyk/crit@latest $ nix profile install github:tomasz-tomczyk/crit # Windows: PS> iwr https://github.com/tomasz-tomczyk/crit/releases/latest/download/crit-windows-amd64.exe -OutFile crit.exe
Then run crit or crit plan.md. Pre-built binaries for ARM64, WSL also available.
Integrates with 11 agents
One-command setup for Claude Code, Cursor, Copilot, Codex, OpenCode, Gemini, Qwen, Hermes, Windsurf, Cline, Aider. Example:
$ claude plugin add tomasz-tomczyk/crit $ claude plugin install crit@crit $ crit install cursor $ crit install github-copilot $ crit install codex
See full setup docs for each agent.
Multi-round and concurrent reviews
Iterate over multiple rounds per file. Run several reviews in parallel without stepping on each other — useful when juggling multiple agents.
Optional sharing and self-hosting
Reviews stay local by default. One click generates a public link — no install, no login for viewers. You can self-host the share target (crit-web) for privacy. Without sharing, reviews auto-delete after 30 days of inactivity.
Small quality-of-life features
- Vim keybindings
- Git diffs (split & unified)
- Round-over-round diffs
- Mermaid diagrams
- Table of contents
- Dark & light themes
- Draft autosave
- Insert suggestions
Who it's for
Developers who frequently review AI-generated plans and code diffs and want a dedicated local UI to batch feedback and iterate without leaving their workflow.
📖 Read the full source: HN AI Agents
👀 See Also
Lucas Gerads demonstrates MCP servers for oscilloscope and SPICE simulator integration with Claude Code
Lucas Gerads built MCP servers for his LeCroy oscilloscope and SPICE simulator, enabling Claude Code to validate SPICE circuits and models, handle embedded programming, and automate data analysis tasks like time axis normalization and data alignment.
Local voice-to-text transcription for OpenClaw using Parakeet TDT 0.6b v3
A developer has converted NVIDIA's Parakeet TDT 0.6b v3 model to run locally via ONNX on CPU, supporting 25 European languages. The model provides an OpenAI-compatible API endpoint through a Docker container, allowing integration with OpenClaw for audio file transcription.
Benchmark shows AI browser automation tools vary 2.6x in token costs despite identical accuracy
A benchmark of 4 CLI browser automation tools using Claude Sonnet 4.6 on 6 real-world tasks found all achieved 100% accuracy, but openbrowser-ai used 36,010 tokens while others used 77,123-94,130 tokens. Tool call count was the strongest predictor of token cost.
OpenRouter Model Pricing and Intelligence-per-Dollar Analysis
A Reddit user compiled OpenRouter API pricing for 16 AI models and calculated intelligence-per-dollar values, identifying MiMo-V2-Flash as best value at $0.09/M tokens and GPT-5.4 as most intelligent at $2.50/M tokens.