Skip to content

Changelog

All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

A large release taking mcpwn from a stdio-only tamper proxy to a multi-transport MCP security tool, while keeping zero runtime dependencies and every transparency/never-block-the-wire invariant intact.

  • Streamable HTTP + SSE transport (mcpwn http --listen <host:port> --upstream <url>): a reverse proxy that runs every JSON-RPC message through the same tamper/correlation/logging pipeline as stdio. Preserves Mcp-Session-Id / Last-Event-ID and SSE event: / id: framing byte-for-byte, relays batches and compressed/opaque bodies unchanged, and forwards the original bytes on any failure. (--mode streamable|sse|auto.) Interactive -i is intentionally disabled for HTTP.
  • Argument tampering for client->server tools/call - new rule actions setArg, rewriteArgs, and fuzzArgs, backed by a tiny zero-dep JSON-path helper (a.b[0].c). fuzzArgs cycles one payload per matching call.
  • Payload library: bundled prompt-injection, sqli, path-traversal, ssrf, and command-injection lists, usable by name in fuzzArgs and via a new mcpwn payloads list | show <name> subcommand.
  • Interception scoping (--scope-method, --scope-tool, --scope-direction, --only-tampered, --scope-affects-log): out-of-scope traffic is always forwarded untouched and hidden from the live log.
  • Rule-file JSON Schema + validation: schema/rules.schema.json plus a hand-rolled validator with precise, collected errors and “did you mean …?” suggestions for mistyped kinds/fields. Invalid rule files exit 2.
  • Baseline / rug-pull detection (--baseline <file>, --baseline-capture <file>, mcpwn baseline capture): diff each tools/list against a captured baseline to catch mid-session tool/schema/description drift.
  • Passive linters + redaction: read-only secret and tool-poisoning detectors, plus --redact / --redact-file to mask secrets in stderr + JSONL.
  • Findings reports (--findings-out <dir>): writes findings.json, findings.md, and findings.sarif on exit.
  • Replay (mcpwn replay <session.jsonl> -- <cmd>): re-drive recorded client->server traffic through a fresh server, with rules applied (--only-method, --repeat).
  • Cross-platform interactive mode: src/engine/tty.ts opens the controlling terminal on POSIX (/dev/tty) and Windows (CONIN$ / CONOUT$), degrading gracefully to auto-forward when headless.
  • Developer tooling (all devDependencies): ESLint + typescript-eslint flat config, Prettier, c8 coverage, and TypeDoc API docs (npm run lint / format / format:check / coverage / docs:api).
  • Architecture: the parse → correlate → rules → intercept → log → forward chain was lifted out of the stdio proxy into a transport-agnostic MessagePipeline + Transport interface, so stdio and HTTP share identical semantics. Correlation is now bidirectional, keyed by (direction, id).
  • test script now discovers all test/**/*.test.js (new test files run automatically); applyAction has a default: throw that is logged (never crashes) for unknown kinds.
  • package.json files now ships schema/ and payloads/ in the tarball.
  • CI builds + unit-tests on a Windows runner in addition to Linux (Node 18/20/22), and gates on lint + coverage.
  • Version bumped to 0.2.0.

First public release.

  • Interactive intercept mode (-i): hold each message and forward / drop / edit ($EDITOR) / quick-tamper, driven from /dev/tty so it never collides with the MCP channel. Falls back to auto-forward when no terminal is present.
  • stdio wrap-and-spawn proxy with transparent pass-through.
  • Live, direction-aware JSON-RPC logging to stderr with request/response correlation.
  • Headless tamper-rule engine: injectToolResult, poisonDescription, matchReplace, and drop actions, loadable from JSON files.
  • JSONL session logging (--log).
  • Bundled vulnerable demo server and one-command end-to-end attack demo.
  • Headless GIF generator (npm run gif) and a vhs tape for the README demo.