Find what your coding agent already spilled.
When you click 'Yes, and don't ask again' in Claude Code (or another coding agent), the full approved command — secrets included — gets saved into a settings file. agent-spill scans your own repos for those saved credentials: settings files, shell history, transcripts, and git history, across every agent folder. It runs entirely on your machine, never uploads anything, and never prints a secret — findings show a masked prefix and a length only.
Requires Node.js 22.13 or newer. There are no runtime dependencies.
npx agent-spill # scan the current directory tree + this machine's agent stores
npx agent-spill --root ~/code # scan a tree of repos (repeatable)
npx agent-spill --json > report.jsonUseful flags:
--root <dir> search under this directory (repeatable; default: cwd)
--no-home skip home-directory agent stores and shell history
--no-history working tree only (no git history)
--ci preset: --no-home --fail-on committed; warns on shallow clones
--fail-on <grade> local | at-risk | committed | pushed (default: at-risk)
--min-confidence <c> high | medium | low (default: medium)
--json machine-readable report on stdout (schema 1)
--verbose expand LOCAL findings and list every location
--max-depth <n> directory walk depth under each root (default: 8)
Exit codes: 0 nothing at or above --fail-on, 1 usage or runtime error, 2 findings at or above --fail-on.
In-repo targets (found under --root, in the work tree and in git history):
| Agent | Paths | Format |
|---|---|---|
| Claude Code | .claude/settings.json, .claude/settings.local.json |
json |
| Claude Code | .mcp.json |
json |
| Cursor | .cursor/mcp.json, .cursor/*.json |
json |
| VS Code / Copilot | .vscode/mcp.json |
json |
| Gemini CLI | .gemini/settings.json, .gemini/policies/*.toml |
json / text |
| Codex | .codex/config.toml, .codex/rules/*.rules |
text |
| SpecStory | .specstory/history/** |
text |
| Continue | .continue/** |
json / text |
| Aider | .aider.chat.history.md, .aider.input.history |
text |
Home-store targets (always graded LOCAL; CLAUDE_CONFIG_DIR and CODEX_HOME are honored):
| Agent | Paths |
|---|---|
| Claude Code | ~/.claude/settings.json, ~/.claude/settings.local.json, ~/.claude.json |
| Claude Code | ~/.claude/history.jsonl, ~/.claude/projects/**/*.jsonl |
| Codex | ~/.codex/rules/*.rules, ~/.codex/config.toml |
| Codex | ~/.codex/history.jsonl, ~/.codex/sessions/**/*.jsonl |
| Gemini CLI | ~/.gemini/settings.json, ~/.gemini/policies/*.toml, ~/.gemini/tmp/*/chats/** |
| Shell | $HISTFILE, ~/.bash_history, ~/.zsh_history, fish history, PowerShell PSReadLine history |
Excluded on purpose: intended credential stores such as ~/.codex/auth.json, ~/.claude/.credentials.json, ~/.gemini/oauth_creds.json, ~/.aws/credentials, ~/.netrc, ~/.npmrc, ~/.docker/config.json, and any .env* file. Flagging the place a secret is supposed to live is noise; other tools cover those.
The worst location wins per secret.
| Grade | Meaning |
|---|---|
PUSHED |
In a commit reachable from refs/remotes/* as of the last fetch. Public or private is not checked. |
COMMITTED |
In local history but not on a remote-tracking ref, or tracked in the index. |
AT-RISK |
In the work tree, untracked, and not ignored by the repository's own rules. |
LOCAL |
Ignored by repository rules, or outside any repo (home stores, shell history). |
historyOnly marks a secret that is in history but no longer in a work-tree file. Deleting the file did not un-leak it.
- Command-embedded passwords:
PGPASSWORD=,MYSQL_PWD=,mysql -p…,sshpass -p,redis-cli -a,--password,curl -u user:pass,Authorization:and API-key headers, credentials in URLs. - Known token formats: GitHub, Anthropic, OpenAI, AWS, Stripe, Slack, Slack webhooks, Google, GitLab, npm, Hugging Face, JWTs (Supabase
service_roleis high;anonis dropped), private-key blocks. - Generic secret-named assignments (
*_TOKEN=,*_SECRET=,*_API_KEY=, …) with a length and entropy floor.
Every captured value goes through a placeholder filter and, for generic rules, an entropy check. JSON and JSONL are parsed and rules run over decoded string values, so PGPASSWORD=\"x\" inside a transcript is seen as the command it was.
- No network calls. Everything is local. There is a test that fails if
src/imports a network module or callsfetch(). - No secret is ever printed. A finding contains only a fixed vendor marker (for example
ghp_is represented bygh), a length, and a per-run fingerprint. The raw value exists only in memory until findings are aggregated. - No disk writes. Fingerprints use a random per-run HMAC key, never persisted. The tests snapshot the temp directory and fixture mtimes around a run.
- Zero runtime dependencies. The supply chain is TypeScript and
@types/node, both dev-only.
# .github/workflows/agent-spill.yml
on: [push, pull_request]
jobs:
agent-spill:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v7
with: { fetch-depth: 0 } # history scan needs full clone
- uses: actions/setup-node@v7
with: { node-version: 26 }
- run: npx --yes agent-spill@0 --ci--ci scans only the checkout, skips home stores, fails on COMMITTED or worse, and warns when the clone is shallow (actions/checkout defaults to depth 1).
Deleting a file does not un-leak a secret. Rotate or revoke the credential at the provider first. Then remove the literal from the agent setting (for Claude settings, replace the literal allow rule with a prefix rule such as Bash(psql:*) and pass the password through the environment). If the value was committed or pushed, purge history with git filter-repo --path <file> --invert-paths (or BFG) and force-push — but forks, clones and caches keep old copies, so rotation is the real fix.
- It does not scan a repository's entire history for every possible secret. Use gitleaks or TruffleHog for that; agent-spill only examines agent-target paths.
- It does not scan other people's repositories or public code.
- It does not validate whether a key is live or contact any provider.
- It does not use a model for classification.
- Cursor's SQLite allowlist (
state.vscdb) and Continue's home-level config are deferred pending verification against real installs.
Apiiro reported on 2026-11-05 that Claude Code's "Yes, and don't ask again" saves the full approved command, secrets included, in plaintext in .claude/settings.local.json. Their public GitHub scan found 104,185 copies, 1,122 (1.1%) with at least one secret; counting history, shared settings and transcripts, 2,539 distinct secrets in 1,994 repositories. 734 secrets appeared only in old git history. Treat those as vendor research: Apiiro sells agent-security products, tested no credentials, contacted no host, and its "real secret" classification rests partly on an LLM reviewer with no third-party replication. The gap that matters here is concrete: the rival scanners miss command-embedded passwords, transcripts, and history where the file was deleted.
Compared with claude-settings-leak-check (Claude settings only; no history, transcripts or other agents) and claudleak (public-repo/org scanning with TruffleHog over full history, inheriting its misses), agent-spill is local-first and targets the command-embedded, transcript, and deleted-file history classes.
Found a real leak? Tell us without pasting any secret value: https://github.com/maxya/agent-spill/issues/new?template=found-a-leak.yml
npm install
npm test # builds src + tests, runs node:test
npm run check # type-check plus testsMIT licensed.