The problem

Claude Code (and every other AI coding CLI) stores your full conversation history as plain-text JSONL on disk — under ~/.claude/projects/ for Claude Code, ~/.codex/sessions/ for OpenAI Codex. Anything you paste — an AWS key, a .env file, a database URL — sits in clear text indefinitely. A typical dev's history accumulates dozens of secrets over months, often without them realizing.

agentsweep scans that history, tells you what leaked, and can redact the secret values in place while preserving the JSONL structure byte-for-byte. It also tells you which keys to rotate, with the right revocation URL for each provider.

Scope of protection: agentsweep itself is fully local and offline — it reads and writes only files on your machine and makes zero network calls. It removes one attack vector: secrets sitting in local history files. It does not affect what your AI provider already received: when you paste a key into Claude Code, Cursor, or any cloud-backed agent, that key already transited the provider's servers before it hit disk. If that concerns you, consider a locally-hosted model (Ollama, LM Studio, OpenCode) where nothing leaves your machine at all — agentsweep pairs especially well with local-model setups.

Why this matters right now

Supply chain attacks are accelerating. In 2024–2025 a wave of malicious npm and PyPI packages — sha256-universal, shailulid, hundreds of typosquats — were caught doing one thing: exfiltrating developer credentials off the machine that installed them. They target environment variables, .env files, shell history, SSH keys, and now AI agent history files.

AI coding assistants have created a new category of credential exposure that didn't exist two years ago:

• You paste a production API key into Claude Code to debug something → it's now in ~/.claude/projects/*/conversations/*.jsonl forever
• A compromised npm package runs postinstall → scans common paths → finds your JSONL history → exfiltrates 50 API keys in one request
• You rotate the key you used in public but forget the dozen others in your history
• Meanwhile your history grows: every .env you asked an AI to help with, every DB URL you shared for debugging, every token you pasted for a one-liner

AI agent history is the new .bash_history — except it contains full context, not just commands. The attack tooling already knows this. agentsweep exists to clean up before it's exploited.

Quick start

pip install uv                  # get uv (skip if you already have it)
uv tool install agentsweep      # install — adds `agentsweep` and `asweep` to PATH
asweep                          # run — interactive menu guides you through everything

That's it. No virtualenv to activate, no PATH fiddling — uv tool install gives the command its own isolated environment.

How it works

agentsweep runs a fixed 5-stage pipeline. scan stops after stage 3; fix continues through redaction.

flowchart LR
    A("🔍 DISCOVER\nwalk history dirs\nstream file list") --> B("⚡ SCAN\nAho-Corasick pre-filter\n191 regex rules + BIP-39")
    B --> C{"secrets\nfound?"}
    C -- "none" --> D("✅ CLEAN\nexit 0")
    C -- "found" --> E("📋 FINDINGS\nshow report\nexit 1")
    E -. "scan only" .-> F("⚠️ ROTATE\nkeys still live")
    E -- "type REDACT" --> G("✏️ REDACT\natomic write · .bak backup\npost-write JSON validation")
    G --> H("🔑 ROTATE\nper-provider\nrevocation links")

    style A fill:#1e3a5f,color:#fff,stroke:#2d5986
    style B fill:#1e3a5f,color:#fff,stroke:#2d5986
    style C fill:#4a3728,color:#fff,stroke:#7a5c3f
    style D fill:#1a4731,color:#fff,stroke:#2d7a52
    style E fill:#4a3a1e,color:#fff,stroke:#7a6030
    style F fill:#4a2020,color:#fff,stroke:#8b3a3a
    style G fill:#1e3a5f,color:#fff,stroke:#2d5986
    style H fill:#2d1e4a,color:#fff,stroke:#5a3a8b

Every write is protected by 8 safety invariants — atomic replace, mandatory .bak backup, symlink rejection, mtime/process gates, and post-write JSONL validation. agentsweep undo restores from backups.

Install

Recommended — isolated, no venv conflicts:

uv tool install agentsweep      # one-time install
uv tool upgrade agentsweep      # update to latest

Try without installing (always runs latest):

uvx agentsweep@latest

Classic pip:

pip install agentsweep
pip install --upgrade agentsweep

Don't have uv? pip install uv or see astral.sh/uv. Requires Python 3.11+.

Usage

Interactive mode

Run with no arguments in a terminal and you get the full experience — banner, numbered menu, typed confirmations before anything destructive, and one-key undo (restores the .bak backups). Any interactive scan that finds secrets ends with an offer to redact them on the spot (type REDACT to confirm):

agentsweep

Scripting is unaffected: any flag, or a piped/redirected stream, skips the menu entirely and behaves exactly as documented below.

Verbs

The legacy flag form agentsweep --fix is still accepted and behaves identically to agentsweep fix.

Source selection

Pick which agent's history to target with --source. The default is claude-code.

agentsweep scan --source claude-code          # ~/.claude/projects/ (default)
agentsweep scan --source codex                # ~/.codex/sessions/
agentsweep scan --source opencode             # OpenCode SQLite store
agentsweep scan --source cursor               # Cursor history
agentsweep scan --source windsurf             # Windsurf history
agentsweep scan --source aider                # ~/.aider/
agentsweep scan --source cline                # Cline history
agentsweep scan --source gemini-cli           # Gemini CLI history
agentsweep scan --source continue-vscode      # Continue (VS Code) history
agentsweep scan --source github-copilot-chat  # GitHub Copilot Chat history
agentsweep scan --source openclaw             # OpenClaw ~/.openclaw/
agentsweep scan --source hermes               # Hermes Agent ~/.hermes/state.db
agentsweep scan --source goose                # Goose ~/.local/share/goose/

Not sure which agents you have installed? list-sources prints every supported source, its history location, and whether that history exists on this machine — so you know which --source values are worth scanning. It reads nothing and writes nothing.

agentsweep list-sources             # all 29 sources + which are on disk
agentsweep list-sources --detected  # only the ones found on this machine
agentsweep list-sources --json      # machine-readable (for scripts/CI)

Override the default root directory to scan any arbitrary folder:

agentsweep scan --root ~/backups/claude-history
agentsweep fix  --root /tmp/history-copy --allow-production

Scan / fix flags

# Machine-readable JSON to stdout (exit 0 = clean, 1 = findings found, 2 = error)
agentsweep scan --json

# Write findings JSON to a file instead of stdout
agentsweep scan --json -o findings.json
agentsweep scan --json --output /tmp/report.json

# Skip .agentsweepignore files
agentsweep scan --no-ignore

Fix-only flags

# Redact Claude Code history in place (--allow-production required for default roots during alpha)
agentsweep fix --allow-production

# Skip creating .bak backups (not recommended)
agentsweep fix --allow-production --no-backup

# Bypass soft safety checks: mtime gate (files modified <60 s ago) and running-process gate
agentsweep fix --allow-production --force

# Combine: non-interactive JSON-mode redaction against a custom root
agentsweep fix --root ~/history-copy --json

Undo flags

# Undo redactions for a specific source
agentsweep undo --source codex

# Undo against a custom root
agentsweep undo --root ~/history-copy

Corruption-prevention guarantees

A redactor that corrupts your history is strictly worse than the leak it's fixing. agentsweep enforces these invariants on every --fix:

1. Redaction happens in parsed JSON, not on raw bytes. Secrets are replaced as string values inside the parsed structure, then re-serialized. Structural damage is impossible by construction.
2. Atomic writes. Every rewrite goes: temp file → fsync() → os.replace() over the original. A crash at any instant leaves either the complete old file or the complete new file — never a torn write.
3. Post-write validation. Before committing, the new content must pass a format-aware check: JSONL — every non-empty line parses as JSON and the line count matches the original; whole-file JSON — the document parses; markdown/plaintext — the line count matches the original; SQLite — the rewritten copy passes PRAGMA integrity_check. If the check fails, the write aborts and the original is untouched.
4. .bak backup by default. Created owner-readable only (mode 0600) since it holds the pre-redaction secrets; refuses to run if a .bak already exists (so prior backups can't be clobbered).
5. Path containment. Refuses any target that doesn't resolve inside one of the source's history trees (e.g. Windsurf's User dir and its ~/.codeium/windsurf/memories/).
6. Symlink rejection. Refuses symlinks outright.
7. mtime window. Refuses files modified in the last 60 seconds (likely an active session). --force overrides.
8. Running-process check. Refuses if a Claude Code process appears to be running. --force overrides.
9. Alpha-stage production gate. --fix against the default ~/.claude/projects/ root requires --allow-production until v1.0.
10. Audit log. Every write appends SHA256 before/after and path to ~/.agentsweep/audit.jsonl.

Recovery

Every redacted file has a sibling *.bak with the original bytes. The easiest way to undo all redactions for a source is:

agentsweep undo                        # undo Claude Code (default)
agentsweep undo --source codex         # undo a specific source
agentsweep undo --root ~/history-copy  # undo against a custom root

If you need to restore a single file manually (e.g. you deleted the undo command's source):

mv session.jsonl.bak session.jsonl

The backups hold the pre-redaction originals — plaintext secrets, so a sweep isn't finished until the leaked keys are rotated and the backups are gone. Once you've rotated:

agentsweep purge                       # delete Claude Code backups (asks first)
agentsweep purge --source windsurf     # delete a specific source's backups
agentsweep purge --yes                 # non-interactive (scripts / CI)

What's detected

191 high-confidence patterns plus a checksum-validated crypto seed-phrase detector — BIP-39 mnemonics (12/15/18/21/24 words; the wallet format behind BTC, ETH, SOL, BNB, ADA, DOGE, LTC, DOT, AVAX and virtually every major chain) and Electrum seeds are confirmed cryptographically (BIP-39 checksum / Electrum version tag), so English prose that happens to use wallet words never false-positives.

The patterns: AWS access keys, GitHub tokens (PAT/OAuth/App/fine-grained), Stripe live/test, OpenAI, Anthropic, Google API, Slack bot/user/webhook, Hugging Face, JWT, PEM private keys, DB URLs with embedded passwords, npm/PyPI/SendGrid/Twilio tokens — plus 167 rules ported from the gitleaks pack covering GitLab, Grafana, HashiCorp Vault/Terraform, DigitalOcean, Shopify, PlanetScale, Databricks, Atlassian, Azure AD, 1Password, Sentry, New Relic, Mailgun, Datadog, Twilio, Twitter/X, Twitch, Yandex, JFrog, Snyk, Mailchimp, curl credentials on the command line, and many more. Patterns are high-precision — false positives are rare, and provider-context rules are keyword-gated so large pastes stay fast.

What's NOT detected

• Custom/proprietary secrets without a recognizable prefix.
• Monero seed phrases (25 words from Monero's own wordlist — planned).
• Unknown tokens that look like arbitrary base64.
• Secrets split across multiple messages.
• Anything inside a binary/non-UTF-8 file.

For deeper detection, run gitleaks or trufflehog alongside agentsweep — their rule packs are more exhaustive. agentsweep's value is the agent-history-specific surface, not the detection engine.

FAQ

Why does uvx agentsweep show an old version? uvx caches tools locally. Use uvx agentsweep@latest to always run the newest version (recommended), or force a cache refresh with uvx --reinstall agentsweep.

Where is OpenCode in the menu? OpenCode support was added in v0.1.1. Run pip install --upgrade agentsweep or uvx agentsweep@latest to get it.

Does agentsweep send my data anywhere? No. It is fully offline — zero network calls during scanning or redacting. The only optional network call is the background update check, which only fetches the latest version number from PyPI.

Contributors

Thanks to everyone who has contributed code, bug reports, and ideas.

License

MIT. See LICENSE.