Skip to content

Security: Harmenszoon/clawback

Security

SECURITY.md

Security & Privacy

clawback sits in the path of your Claude Code traffic and writes an exhaustive log of every request and response. Please read this before running it anywhere other than your own machine.

Threat model / intended deployment

  • Single-user, localhost only. The proxy binds to localhost:3456 by default and has no authentication. The proxy itself holds and adds no credentials — it forwards whatever Authorization / x-api-key headers each client sends — so an attacker reaching the socket cannot spend your quota. What they can do is relay their own traffic through your machine (hiding their origin) and grow your logs without bound. Do not bind it to a public interface (PROXY_HOST) or expose the port.
  • It forwards your real API credentials upstream. The Authorization / x-api-key headers your client sends are passed through to api.anthropic.com unchanged on every request (and redacted in the logs).
  • Request bodies are unbounded (client_max_size=0) by design, which is fine for a local single client but unsafe on an exposed interface.

The logs contain sensitive data

logs/ holds your prompts, file contents, command output, and tool results in full. Treat the directory as sensitive:

  • It is gitignored and must never be committed.
  • Secrets-bearing headers (Authorization, x-api-key, Cookie, Set-Cookie, Proxy-Authorization, X-Claude-Code-Session-Id) and the request-body metadata.user_id telemetry blob are redacted in the log copy — but the conversation content is not. Redaction makes logs tidier, not safe to publish.
  • Review and sanitize before sharing a log for a bug report.

Reporting a vulnerability

This is a small, public-domain, single-author project. If you find a security issue, please open a GitHub Security Advisory (preferred) or a regular issue if it is not sensitive. There is no formal SLA, but reports are welcome and will be addressed on a best-effort basis.

There aren't any published security advisories