Skip to content

mesa-dot-dev/saguaro

npm version CI License

Saguaro Local Review

Local code review for coding agents. Rules enforced inside Claude Code, Codex, and Cursor. Your agent fixes its own mistakes while context is hot. Free. Apache 2.0.

Install

npm (macOS + Linux, requires Node.js 20+)

npm install -g @mesadev/saguaro

Or run without installing:

npx @mesadev/saguaro review

Homebrew (macOS)

brew install mesa-dot-dev/homebrew-tap/saguaro

Shell script (macOS + Linux)

curl -fsSL https://raw.githubusercontent.com/mesa-dot-dev/saguaro/main/install.sh | sh

Saguaro uses your existing agent CLI subscription (Claude Code, Codex, Gemini) when available, no API key needed. If no agent CLI is installed, set ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY in your environment.

Quickstart

# 1. Initialize Saguaro in your repo
sag init

# 2. Generate rules from your codebase (or use the starter rules)
sag rules generate

# 3. Run a review
sag review

sag init walks you through setup:

  • Creates .saguaro/config.yaml and .saguaro/rules/
  • Offers three options: generate rules from your codebase, use starter rules, or start from scratch
  • Sets up Claude Code integration automatically (MCP server, hooks, slash commands)
  • Stores your API key in .env.local if provided

How It Works

  1. Match rules to files — Rules load from .saguaro/rules/ as markdown with YAML frontmatter. Changed files match against rule globs using minimatch. Rules without globs apply to all changed files.
  2. Gather context — Saguaro builds an import graph (tree-sitter + SWC) showing the "blast radius" — files that depend on your changes and files your changes depend on. Context is token-budgeted so reviews stay fast.
  3. AI review per file — Changed files batch together (configurable via files_per_batch). An AI agent checks each batch against matched rules with the diff and context. If a rule is violated, Saguaro prints it. If nothing is wrong, silence.

Violations exit 1. Clean reviews exit 0.

Claude Code Integration

sag init sets up three things automatically:

Hooks

Two hooks install into .claude/settings.json:

  • PreToolUse — Fires before Edit and Write tool calls. Injects relevant rules so Claude knows your team's conventions before writing code.
  • Stop — Fires when Claude finishes a turn. Saguaro diffs uncommitted changes against the base branch and reviews them. Violations block Claude and ask it to fix before completing.
sag hook install    # Enable (done by sag init)
sag hook uninstall  # Disable

MCP Server

Saguaro registers as an MCP server via .mcp.json. Claude Code discovers it on startup and gains access to Saguaro's tools (review, create rules, generate rules).

Slash Commands

Command What it does
/saguaro-review Run a code review manually
/saguaro-createrule Create a new rule with AI assistance
/saguaro-generaterules Auto-generate rules from your codebase

Other Agents

Saguaro works with any coding agent. The CLI is agent-agnostic.

Agent Integration
Codex CLI Run sag review manually or in CI
Gemini CLI Same CLI and rules
Cursor JSON output with deeplinks to violations (--output json, cursor_deeplink: true in config)
CI pipelines sag review --base origin/main — exits 1 on violations

Background Daemon

For long-running agent sessions, the daemon runs classic senior engineer level reviews asynchronously in the background. Findings are advisory and surfaced on the next agent turn — independent of the rules review system.

sag daemon start    # Start the review daemon
sag daemon stop     # Stop it
sag daemon status   # Check if the daemon is running

The daemon runs an HTTP server with a SQLite-backed job queue and worker pool. The stop hook posts diffs to the daemon instead of running reviews inline. Workers claim jobs, spawn AI agents, and store results. The hook client polls for completion.

Enable in .saguaro/config.yaml:

daemon:
  enabled: true

Rules

Rules live in .saguaro/rules/ as markdown files with YAML frontmatter. Each rule defines what to check, which files it applies to, and how severe a violation is.

---
id: no-console-log
title: No console.log in production code
severity: warning
globs:
  - "**/*.ts"
  - "**/*.tsx"
  - "!**/*.test.ts"
---

console.log, console.warn, and console.debug should not appear in
production code. Use a structured logging library instead.

### Violations

```
console.log("Processing order", orderId, customerEmail);
```

### Compliant

```
logger.info("Processing order", { orderId });
```

Fields:

Field Required Description
id Yes Unique identifier, kebab-case
title Yes Human-readable name
severity Yes error (blocks, exit 1), warning (logged), or info
globs No File patterns to match. Default: all files. Prefix ! to exclude.

The markdown body is the instruction the AI uses to evaluate the rule. Include ### Violations and ### Compliant sections with code examples.

Saguaro reviews any text file — rules are language-agnostic. The import graph supports TypeScript, JavaScript, Python, Go, Rust, Java, and Kotlin.

For a deep dive on writing effective rules, see Writing Rules.

CLI Reference

Command Description
sag init Set up Saguaro in your repo (config, rules, hooks, integrations)
sag review Review code changes against your rules
sag rules generate Auto-generate rules by analyzing your codebase
sag rules create [dir] Create a new rule with AI assistance, scoped to a directory
sag rules list List all rules with IDs, titles, and severity
sag rules explain <id> Show full details for a rule
sag rules validate Check all rule files for correct structure
sag rules delete <id> Delete a rule
sag rules locate Print the path to the rules directory
sag index Build the import graph for richer review context
sag hook install Enable automatic reviews in Claude Code
sag hook uninstall Disable automatic reviews
sag daemon start Start the background review daemon
sag daemon stop Stop the background review daemon
sag daemon status Check if the daemon is running
sag model Switch AI provider and model interactively
sag stats Show review history and cost analytics

sag review Options

-m, --mode     Review mode: rules, classic, or full       [default: "rules"]
-b, --base     Base branch to diff against                [default: "main"]
    --head     Head ref to review                         [default: "HEAD"]
-o, --output   Output format: console, json               [default: "console"]
-v, --verbose  Show detailed progress                     [default: false]
    --debug    Write debug logs to .saguaro/.tmp/         [default: false]
-c, --config   Path to config file                        [default: ".saguaro/config.yaml"]
    --rules    Path to rules directory                    [default: ".saguaro/rules/"]

sag rules create Options

sag rules create [target]

    target         Directory the rule targets (e.g. src/api)
    --intent       What the rule should enforce
    --severity     error, warning, or info
    --title        Rule title (auto-generated if omitted)
    --skip-preview Skip the file-match preview step

Configuration

.saguaro/config.yaml is created by sag init:

# AI model for reviews
model:
  provider: anthropic       # anthropic | openai | google
  name: sonnet              # CLI alias (e.g. "sonnet", "opus") or model ID

# Output settings
output:
  cursor_deeplink: true      # Print Cursor IDE links for violations

# Review tuning
review:
  max_steps: 10              # Max tool-calling steps per review batch
  files_per_batch: 2         # Files reviewed together per batch

# Hook settings
hook:
  enabled: true              # Master switch for all Saguaro hooks
  stop:
    enabled: true            # Rules review after each code change

# Background daemon (classic reviews)
# daemon:
#   enabled: true

API keys are loaded from environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY) or from .env.local / .env files.

Pre-push Hook

#!/bin/bash
# .git/hooks/pre-push
sag review --base origin/main

FAQ

How is this different from CodeRabbit / Greptile / etc?

Those tools are AI reviewers that comment on your PRs. Saguaro Local Review is a rules engine. You define what matters to your team, Saguaro enforces it during development — not after the code is already in a PR. No noise, no generic suggestions. If nothing is violated, you hear nothing.

How much does it cost?

You use your own API key. Cost depends on your model choice, codebase size, number of rules, and how often you run reviews. sag stats shows your usage history.

Can I use it without Claude Code?

Yes. The CLI works standalone. Run sag review from any terminal or CI pipeline. The Claude Code integration (hooks, MCP, slash commands) is optional.

Where does my data go?

Nowhere. Saguaro runs locally. Your code is sent to the AI provider you configure (Anthropic, OpenAI, Google) for review. Nothing touches Saguaro's servers.

What languages are supported?

Saguaro reviews any text file — rules are language-agnostic. The import graph supports TypeScript, JavaScript, Python, Go, Rust, Java, and Kotlin.

Can I use it in CI?

Yes. sag review --base origin/main exits 1 if violations are found. Use --output json for structured output.

What AI providers are supported?

Anthropic (Claude), OpenAI (GPT-4o, o3), and Google (Gemini).

What's the background daemon?

An optional async review system for long-running agent sessions. Runs classic (senior-engineer-style) reviews in parallel without blocking your agent. See Background Daemon.

License

Apache-2.0.

About

No description, website, or topics provided.

Resources

License

Code of conduct

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages