We build Claude Code plugins that find real bugs in real runtimes.
Not lint. Not style warnings. Bugs — the kind that require understanding how the runtime actually works: reference count errors, extraction boundary violations, thread-safety assumptions that break under free-threading, C API misuse that corrupts interpreter state.
Every toolkit is anchored to a confirmed historical bug. Every finding requires a runnable reproducer before it's considered a finding.
| Toolkit | What it reviews | What it finds |
|---|---|---|
| cext-review-toolkit ⭐ 28 | CPython C extensions | API misuse, refcount errors, type slot bugs, ABI breaks |
| cpython-review-toolkit ⭐ 10 | CPython C source | Null safety, error paths, memory management, interpreter bugs |
| cpython-security-toolkit | CPython stdlib (Lib/) |
Archive escapes, validation gaps, resource amplification, audit hook bypass |
| ft-review-toolkit | CPython C extensions | Free-threading races, GIL assumptions, ThreadSanitizer triage |
| pypy-review-toolkit | PyPy / RPython | RPython contract violations, JIT-visible correctness bugs |
| rustpy-review-toolkit | RustPython (Rust source) | Python-reachable panics, GC traverse gaps, object model soundness |
| code-review-toolkit | Any codebase | General correctness, quality, architecture |
Findings in the wild: pypy-review-findings — documented bugs in PyPy's own RPython implementation, confirmed and reproduced.
All toolkits run as Claude Code plugins. Point one at a source tree and it dispatches specialized agents that understand the runtime's specific invariants — not just syntax, but the semantics that matter for that particular codebase.
The flow is always the same:
named invariant → coverage analysis across all code paths
→ reproducer generation
→ confirmed finding
A candidate without a reproducer stays a candidate. This is the constraint that makes the output useful to upstream maintainers rather than just noisy.
git clone https://github.com/ReviewToolkits/<toolkit-name>
cd <toolkit-name>
claude --plugin-dir plugins/<toolkit-name>Requires Claude Code and Python 3.11+.
Toolkits that find security issues (cpython-security-toolkit) route findings to security@python.org — never the public tracker. Each security toolkit includes a WORKING_WITH_MAINTAINERS.md that covers the disclosure process in full.
Built by @devdanzin and @BHUVANSH855.
