Sift reviews your code before it leaves your machine. It filters noise from diffs, maps the blast radius of your changes, and gives you a plain-English summary — all in your terminal.

What You Get

1. The Diff — A noise-collapsed view showing only meaningful changes. Whitespace, import reordering, formatting, and comment-only changes are filtered out.

2. The Flow — Call graph visualization showing which files are affected downstream. Built using lightweight static analysis for TypeScript and Python.

3. The Why — Three lines of plain-English explanation: Problem, Fix, Risk. Max 25 words per line.

Example Output

Summary
  Problem: Auth middleware stores session tokens in plaintext cookies
  Fix:     Switched to httpOnly signed cookies with 24h expiry
  Risk:    Existing sessions will be invalidated on deploy

Blast radius: 3 files changed, 2 affected downstream

  ┌─────────────────────────────────────────────────┐
  │ src/middleware/auth.ts (setSession) CHANGED      │
  └──────────────────────┬──────────────────────────┘
                         │
                         v
          ┌──────────────────────────────┐
          │ src/routes/api.ts → handler  │
          └──────────────────────────────┘

2 issues found:
  ! src/middleware/auth.ts:42 Cookie maxAge is set in ms but comment says seconds
  ! src/middleware/auth.ts:58 No CSRF token validation on session refresh endpoint

Confidence: 82%  |  +34 -12  |  8 noise lines filtered
Tokens: 1,247 in / 312 out  |  Est. cost: $0.0085

Quick Start

Claude Code (MCP Server)

Install sift and register it as an MCP server in one step:

npm install -g sift-review
claude mcp add sift -- sift-mcp

Restart Claude Code, then just say "review my changes". Sift auto-discovers your repo and runs the review. It also runs automatically before any git push.

Without a global install: claude mcp add sift -- npx -y sift-review sift-mcp

CLI

# Review changes on your current branch
sift review

# Review against a specific base
sift review --base main

# Specify both base and head
sift review --base main --head feature-branch

# Run without installing
npx sift-review review

Requirements

• Node.js >= 20
• Git
• ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN environment variable set

MCP Tools

When running as an MCP server, sift exposes these tools to Claude Code:

| Tool | Description | |------|-------------| | review_diff | Run a full AI code review (auto-discovers repos) | | get_blast_radius | Get the call graph showing downstream impact | | get_file_details | Drill into a specific file's changes | | get_diff_page | Page through the diff one file at a time | | discover_repos | Scan a directory for git repos with changes |

Configuration

Create an optional .sift.json at your project root:

{
  "model": "claude-sonnet-4-6-20250514",
  "max_diff_lines": 500,
  "base_branch": "main",
  "max_issues": 10
}

| Field | Default | Description | |-------|---------|-------------| | model | claude-sonnet-4-6-20250514 | Claude model for analysis | | max_diff_lines | 500 | Max diff lines sent to Claude | | base_branch | main | Default base branch for diffing | | max_issues | 10 | Max issues to return | | api_key_env | ANTHROPIC_API_KEY | Environment variable name for the API key |

Proxy / LiteLLM Support

Sift works with LiteLLM and other Anthropic API-compatible proxies. Set the standard environment variables:

export ANTHROPIC_BASE_URL=https://your-proxy.example.com/
export ANTHROPIC_AUTH_TOKEN=your-token

Then configure the model name in .sift.json:

{
  "model": "your-proxy-model-name"
}

The Anthropic SDK reads these environment variables automatically — no additional sift configuration needed. Cost estimation is disabled when using a proxy.

How It Works

1. You run sift review (or it runs via a pre-push hook)
2. Sift extracts the diff between your branch and the base
3. Noise is filtered out — whitespace, import reordering, formatting, comments
4. The filtered diff is sent to Claude for review
5. Static analysis builds a call graph showing affected downstream files
6. AI review and graph building run in parallel
7. Results are formatted and displayed in your terminal

Bypass

git push --no-verify

License

MIT