Argus

Your coding agent shouldn't grade its own homework.

Argus is a local-first AI code review platform for teams that want an independent reviewer, not a model grading its own output. It combines structural analysis, semantic search, git history intelligence, and LLM-powered review so comments are grounded in the full codebase instead of just the patch.

The short version is simple: your coding agent should not grade its own homework.

git diff HEAD~1 | npx argus-ai review --repo .

Live page: merup.me/argus

Why Argus?

• Independent review — your AI agent wrote the code, a different AI reviews it. No self-grading.
• Full codebase context — reviews use structural maps, semantic search, git history, and cross-file analysis. Not just the diff.
• Zero lock-in — works with OpenAI, Anthropic, or Gemini. Switch providers in one line. Gemini free tier = zero cost.
• One binary, nine tools — map, diff, search, history, review, describe, feedback, doctor, MCP server. Composable Unix-style subcommands.

Get Started in 60 Seconds

# 1. Install via npm
npx argus-ai init          # creates .argus.toml

# 2. Set your key (Gemini, Anthropic, or OpenAI)
export GEMINI_API_KEY="your-key"

# 3. Review your changes
git diff HEAD~1 | npx argus-ai review --repo .

Install

Homebrew (macOS/Linux)

brew tap Meru143/argus
brew install argus

npm (Recommended)

npx argus-ai --help
# or
npm install -g argus-ai

Cargo

cargo install argus-ai

From Source

cargo install --path .

Subcommands

review — AI Code Review

Run a context-aware review on any diff or PR.

# Review local changes
git diff main | argus review --repo .

# Review a GitHub PR (posts comments back to GitHub)
argus review --pr owner/repo#42 --post-comments

describe — PR Descriptions

Generate structured, conventional-commit PR descriptions from your changes.

# Generate description for staged changes
argus describe

# Generate for a specific PR
argus describe --pr owner/repo#42

feedback — Improve Reviews

Rate comments from your last review to train Argus on your preferences.

# Start interactive feedback session
argus feedback

map — Codebase Structure

Generate a ranked map of your codebase structure (tree-sitter + PageRank).

argus map --path . --max-tokens 2048

search — Semantic Search

Hybrid code search using embeddings (Voyage/Gemini/OpenAI) + keywords.

argus search "auth middleware" --path . --limit 5

history — Git Intelligence

Detect hotspots, temporal coupling, and bus factor risks.

argus history --path . --analysis hotspots --since 90

diff — Risk Scoring

Analyze diffs for risk based on size, complexity, and diffusion.

git diff | argus diff

mcp — MCP Server

Connect Argus to Cursor, Windsurf, or Claude Code.

argus mcp --path /absolute/path/to/repo

doctor — Diagnostics

Check your environment, API keys, and configuration.

argus doctor

GitHub Action

Add automated reviews to your PRs:

name: Argus Review
on: [pull_request]
permissions:
  pull-requests: write
  contents: read
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - name: Install Argus
        run: npm install -g argus-ai
      - name: Run Review
        env:
          GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          argus-ai review \
            --diff origin/${{ github.base_ref }}..HEAD \
            --pr ${{ github.repository }}#${{ github.event.pull_request.number }} \
            --post-comments \
            --fail-on bug

MCP Setup

Add to ~/.mcp.json or project .mcp.json:

{
  "mcpServers": {
    "argus": {
      "command": "argus",
      "args": ["mcp", "--path", "/absolute/path/to/repo"]
    }
  }
}

Add to generic MCP settings:

{
  "argus": {
    "command": "argus",
    "args": ["mcp", "--path", "."]
  }
}

Configuration

Run argus init to generate a .argus.toml:

[review]
# max_comments = 5
# min_confidence = 90
# skip_patterns = ["*.lock", "*.min.js", "vendor/**"]

Custom Rules

Argus supports natural language custom rules. Create a file at .argus/rules.md (or .argus/rules/**/*.md) to guide the AI reviewer.

Example .argus/rules.md:

- Always suggest using `anyhow::Result` instead of `Result<T, Box<dyn Error>>`.
- Flag usage of `unwrap()` in production code; suggest `expect()` or error handling.
- Ensure all public functions have doc comments.

LLM Providers

For custom provider endpoints, set [llm].base_url. OpenAI-compatible and Gemini
base URLs can be provided either with or without the version suffix, for example
https://openrouter.ai/api or https://openrouter.ai/api/v1.

Embedding Providers

Zero-cost setup: Use Gemini for both LLM and embeddings with a free API key.

[llm]
provider = "gemini"

[embedding]
provider = "gemini"

Environment Variables

Architecture

                    ┌─────────────┐
                    │   argus     │
                    └──────┬──────┘
                           │
          ┌────────────────┼────────────────┐
          ▼                ▼                ▼
  ┌───────────────┐ ┌───────────┐ ┌──────────────┐
  │ argus-review  │ │ argus-mcp │ │  subcommands │
  └───────┬───────┘ └─────┬─────┘ └───────┬──────┘
          │               │               │
    ┌─────┴─────┬─────────┘               │
    ▼           ▼           ▼             ▼
┌─────────┐ ┌─────────┐ ┌──────────┐ ┌──────────┐
│ repomap │ │difflens │ │ codelens │ │ gitpulse │
└─────────┘ └─────────┘ └──────────┘ └──────────┘

Crate dependency order:

core (no internal deps)
  ├── repomap (core)
  ├── difflens (core)
  ├── gitpulse (core)
  ├── codelens (core, repomap)
  └── review (core, repomap, difflens, codelens, gitpulse)
        └── mcp (core, review)

Contributing

See CONTRIBUTING.md.

License

MIT