Leakferret

MCP-native secret scanner — verified findings, agent-applied rewrites.

leakferret is one fast Rust binary that is engine, CLI, and MCP server. It finds hardcoded secrets in your code, calls the provider to confirm which ones are actually live, and rewrites the leak in place to read from an environment variable. It runs in your terminal, in CI, and as a tool your coding agent calls before it commits — and the raw secret never leaves your machine.

What it looks like

Say you accidentally commit a real key, plus the usual noise:

# .env  — every key below is fabricated for this example
STRIPE_SECRET_KEY=sk_live_FAKE_example_not_a_real_key   # fabricated
GITHUB_TOKEN=ghp_FAKE_example_not_a_real_token          # fabricated
SENDGRID_API_KEY=${SENDGRID_API_KEY}                    # a reference — not a leak
ADMIN_PASSWORD=changeme                                 # a placeholder — not a leak
AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE                  # AWS's public docs example

leakferret verify calls each provider and tells you what is real and live — and stays quiet on the rest:

$ leakferret verify .
.env
  L2  UNKNOWN    CRITICAL  stripe_secret   sk_l..._key
  L3  VERIFIED   CRITICAL  github_token    ghp_...oken   ← live, rotate it now

2 findings · 1 verified · 1 unknown

_{The keys above are fabricated, so the VERIFIED line illustrates what a genuinely live key reports — on these examples both would be UNKNOWN.}

The ${SENDGRID_API_KEY} reference, the changeme placeholder, and the well-known AKIAIOSFODNN7EXAMPLE example are recognized and left out — that precision is the point. Then leakferret rewrite --apply rewrites the hardcoded key in your code (it leaves .env files alone — there's nothing sensible to rewrite a secret to there):

  # app/billing.rb  (fabricated example)
- Stripe.api_key = "sk_live_FAKE_example_not_a_real_key"   # fabricated
+ Stripe.api_key = ENV.fetch("STRIPE_API_KEY")

…and appends STRIPE_API_KEY= to .env.example with a seed command for your secret manager. Find → confirm live → fix, with almost no false alarms.

The full secret value never leaves your machine. Only a redacted AKIA...4XYZ preview is ever written to a report, log, or network message.

Quick start

Install however you like — every package ships the same prebuilt binary.

# Ruby gem
gem install leakferret

# npm (CLI)
npm i -g @leakferret/cli

# Go
go install github.com/leakferrethq/leakferret-go/cmd/leakferret@latest

# Native binary — download from GitHub Releases, unpack, and put it on $PATH:
#   https://github.com/leakferrethq/leakferret/releases

# Rust, from source
cargo install leakferret-cli

Then scan the current directory:

leakferret scan .

scan respects .gitignore and also reads dotfiles such as .env. Add --git to walk commit history instead of the working tree.

Every wrapper honors a LEAKFERRET_BIN environment variable pointing at a local binary, for offline or development use.

How it works

leakferret runs findings through a five-station pipeline. Each station only sees what it needs, and the raw secret never advances past disk.

1. Scan — a fast regex pre-filter over your files, across 60+ secret types. Respects .gitignore, reads dotfiles like .env, and (with --git) walks history.
2. Catalog — every candidate is checked against a signed database of known-public example credentials: Stripe test keys, AKIAIOSFODNN7EXAMPLE, jwt.io samples, RFC examples. Matches are marked FIXTURE so documented examples never raise a false alarm. The catalog is bundled with the binary and can be refreshed and signature-verified.
3. Classify — each remaining candidate gets a verdict: REAL, FIXTURE, or UNKNOWN. This runs offline by default (path rules plus dummy-marker heuristics), or asks the host editor or agent's own language model — no extra API key, no added cost.
4. Verify — makes a single harmless API call to the provider to confirm a key is LIVE. Around 25 providers are covered natively (AWS SigV4, GitHub, GitLab, Stripe, OpenAI, Anthropic, Slack, Twilio, SendGrid, Mailgun, Datadog, Heroku, npm, PyPI, DigitalOcean, Hugging Face, Groq, Replicate, Notion, Postman, Figma, Linear, Square, Shopify, Databricks), with a trufflehog binary fallback for the rest. The call goes straight from your machine to the provider — leakferret has no servers.
5. Rewrite — swaps a hardcoded literal for an environment-variable lookup (ENV.fetch / os.environ / process.env), appends a line to .env.example, and prints seed commands for your secret manager (env, Vault, Doppler, AWS Secrets Manager, or Infisical).

A baseline stores one-way HMAC fingerprints of known findings — never the raw secret — so CI can fail only on new leaks.

Use it with AI agents (MCP)

MCP (Model Context Protocol) is the open standard for giving coding agents tools. Agents hardcode secrets too, and nobody reviews their diffs line by line — leakferret lets the agent self-check before it commits.

Start the server over JSON-RPC on stdio:

npx @leakferret/mcp

Add it to your mcpServers config (Claude Desktop, Cursor, Continue, Claude Code):

{
  "mcpServers": {
    "leakferret": {
      "command": "npx",
      "args": ["@leakferret/mcp"]
    }
  }
}

For Claude Code, save that block as .mcp.json in your project root, or add it with one command:

claude mcp add leakferret -- npx -y @leakferret/mcp

If you installed the native binary, you can point at it directly instead:

{
  "mcpServers": {
    "leakferret": {
      "command": "leakferret",
      "args": ["mcp"]
    }
  }
}

leakferret is also listed in the MCP Registry as io.github.leakferrethq/leakferret, so registry-aware clients can discover it.

Tools exposed: scan_repository, classify_candidates, verify_finding, propose_rewrite, and baseline_diff. A classify prompt is also provided so an agent can classify candidates inline using the model it already has. Two read-only resources expose the engine's catalog as context: leakferret://secret-types (every detectable pattern) and leakferret://verifiers (the live-verification providers).

How it compares

gitleaks is the fastest pre-commit regex blocker. trufflehog set the bar for live verification. detect-secrets owns the baseline-a-legacy-repo workflow. GitGuardian is the paid platform with the broadest detectors and a dashboard. leakferret does the regex pre-filter, verifies which keys are live, rewrites the leak to an env var, and runs as an MCP server so coding agents check their own diffs. Full writeup: leakferret.com/compare.

CLI reference

leakferret scan      Regex pre-filter only (no classifier, no verifier)
leakferret verify    Scan + classify + provider verification
leakferret rewrite   Scan + classify + propose/apply ENV-fetch rewrites
leakferret org       Scan every public repo owned by a GitHub user/org
leakferret baseline  Manage the per-repo fingerprint baseline
leakferret catalog   Load and inspect the fixture catalog
leakferret mcp       Start the MCP server on stdio

Scan a whole GitHub account or org in one go (handy for finding leaks across all your public repos before someone else does):

leakferret org leakferrethq            # all public repos for that owner
leakferret org myco --token "$GITHUB_TOKEN" --format sarif > leaks.sarif

Common flags:

# scan
leakferret scan .                              # working tree
leakferret scan . --git                        # scan HEAD's commit history
leakferret scan . --git --all                  # scan every branch / tag
leakferret scan . --git --since HEAD~50        # bounded history window