DiffGuard

Audit code changes for security issues and unintended behavior changes using your own Anthropic API key. DiffGuard reads a git diff, sends it to Claude for review, and prints structured findings — or exits cleanly when nothing's wrong. Built for the terminal: works on any machine with git and Node, drops straight into CI and pre-commit hooks.

Install

npm install -g @figo5/diffguard

Setup

diffguard config

diffguard config walks you through choosing a provider (Anthropic, OpenAI, OpenRouter, Groq, Ollama, or a custom OpenAI-compatible base URL), picking a model, and storing your API key in the platform keychain. Afterwards, plain diffguard just works — no env vars needed. Non-interactive setup is available too:

diffguard config --provider=ollama --model=gemma4:e2b
diffguard config --show    # print the effective config (never the key)
diffguard config --reset   # clear the saved config

Note: diffguard config --key=... puts the key on the command line, where it's visible in shell history and to anyone who can list processes on the machine. Prefer the interactive wizard (diffguard config) or diffguard key set, which read the key from a hidden prompt or stdin instead.

You can still configure everything with env vars instead — see the Configuration table. Env vars always override the saved config file.

Usage

diffguard            # audit the working-tree diff
diffguard --staged   # audit the staged diff (pre-commit)
diffguard --json     # machine-readable output
diffguard config    # one-time setup: provider, model, and key

In coding agents

DiffGuard is available as an on-demand audit tool inside coding agents:

• Pi — install the extension at ~/.pi/agent/extensions/pi.ts (or .pi/extensions/pi.ts).
• MCP-capable agents (Claude Code, Codex, Cursor, Windsurf, Copilot, Cline) — run the diffguard-mcp stdio server and add it as an MCP server.

The agent calls the diffguard_audit tool with a scope argument:

The tool returns a text report plus structured findings. The API key is resolved from the agent's own credential (Pi), ANTHROPIC_API_KEY, or the platform keychain (diffguard key set). The key never enters the agent's context.

The agent layer reads the same saved config file as the CLI, so the provider and model you configure with diffguard config apply to Pi and MCP too. (Local providers like Ollama still need a key in the agent's environment — set ANTHROPIC_API_KEY to any value.)

Exit codes

Use it as a pre-commit gate:

# .git/hooks/pre-commit, or via a pre-commit framework
diffguard --staged || exit 1

Example

$ diffguard
[security] high  src/auth.ts:42
  Credentials compared with == instead of a constant-time comparison.
  Fix: use crypto.timingSafeEqual.

2 findings (1 security, 1 behavior)

Configuration

Precedence: env var > config file > default. The API key is never stored in the config file — it lives in the platform keychain (diffguard key set / diffguard config) or the ANTHROPIC_API_KEY env var. The CLI checks the env var first, then the keychain.

How it works

1. DiffGuard collects the git diff and some lightweight repo context.
2. It builds a prompt instructing Claude to return findings as strict JSON, and streams the response (cancel with Ctrl-C).
3. The response is parsed and validated — malformed findings are dropped and counted, never silently accepted.
4. Findings print as a grouped report, or as raw JSON with --json.

If the diff is empty, DiffGuard exits 0 without calling the API. If the diff is too large, it's truncated to DIFFGUARD_MAX_DIFF_TOKENS and the audit is marked partial.

Privacy

• The diff you audit is sent to Anthropic to generate findings. Don't audit code you aren't comfortable sending to a third-party API.
• DiffGuard makes no other network calls and collects no telemetry.

Roadmap

• Pre-commit hook installer (diffguard install).
• --path filtering, --fail-on <severity>.
• Test-gap and license-compliance checks alongside security/behavior findings.

License

MIT — see LICENSE.