culprit

Root-cause analysis for agent-driven development.
Your coding agent finds what introduced a bug, and verifies its own fix is complete before it commits.

Quick start · Tools · Accuracy · Architecture · MCP registry

A real run. The fix patched one caller of parse_config and missed two, so it scores partial.

Why an agent needs this

An agent can dig through git history to find what broke, and it is good at it. What it will not reliably do is spend a dozen tool calls on that every time, or stop and check whether the helper it just changed has three other callers before it commits. culprit turns both into one deterministic call that returns structured JSON it can act on:

• ⚡️ Instant. No test runs, no bisect. Benchmarked at 41% fewer agent tool calls.
• 🔒 Read-only and offline. Never writes to your repo or PR. No API key, no network.
• 🔌 MCP-native. 11 tools over stdio, on the official MCP registry, plus a one-command plugin.
• 🗣️ Language-agnostic. Suspects work anywhere git does. Blast radius reads 12 language families.

Does it actually work?

Benchmarked against 50 real regressions, 25 from git and 25 from systemd, where the introducing commit is known from each fix's Fixes: trailer (author-verified ground truth). Given only the fix commit, culprit blames the removed lines to rank the commits that introduced the bug.

Deterministic and offline, on large C codebases the engine has never seen. Reproduce with python benchmarks/run.py, which clones the repos and scores every case.

But an agent can run git blame itself

It can, and it is good at it. So the honest question is not whether culprit is accurate, but whether it beats or helps an agent that has git and no culprit. Same 10 regressions, three arms, isolated so no arm can read the answer from the fix commit:

The agent wins on accuracy, and culprit does not make it more accurate. What culprit does is get it there in roughly half the steps for about 8,000 fewer tokens per investigation. Tool calls dropped in 10 of 10 cases, tokens in 7 of 10. On the hardest case the agent needed 42 tool calls and 124k tokens alone, against 16 calls and 84k with culprit.

Those savings are already net of reading culprit's output, which is counted in the third column. culprit's own run is one deterministic call, no model, under a second.

So culprit is an accelerator for an agent, not a replacement for one, and not a smarter blamer than one. Method, caveats (n=10, and this slice is easier than the full 50), and reproduction: benchmarks/agent/.

Quick start

Claude Code plugin (installs the MCP server plus a skill that tells the agent when to use it):

/plugin marketplace add noordeen123/culprit
/plugin install culprit@culprit

Any MCP client (Cursor, Windsurf, VS Code, Codex CLI, Zed, Continue, Cline, Amazon Q, Goose):

{
  "mcpServers": {
    "culprit": {
      "command": "uvx",
      "args": ["--from", "culprit[mcp]", "culprit-mcp"]
    }
  }
}

CLI, no agent required:

uvx culprit                     # run from PyPI on demand
rca --verify-fix patch.diff     # exit 0 if complete, 1 otherwise

Needs Python 3.10+ and uv for the MCP server. The CLI runs on 3.9+.

The tools

verify_fix, the pre-commit gate

The agent runs this on its own diff before committing:

• verdict: complete when no untouched call site is left behind, partial when one was missed, risky when risk is high. The verdict is the completeness axis; test coverage is the separate confidence axis on risk_level.
• untouched_references: the exact files that still use the changed symbol. This is the list the agent goes and patches.
• skipped_symbols, adds_test, notes: what was too widely used to check, whether a test shipped, and what to do next.

Loop until complete, then commit. That is the whole idea.

CLI and CI

rca                            # current branch vs the configured base
rca --last                     # the latest commit only
rca --pr 16786                 # a specific PR (uses the PR's own base)
rca --trace crash.txt          # RCA from a stack trace, no fix or PR needed
rca --verify-fix patch.diff    # check a diff before committing
rca --select-tests             # print the tests to run for this change
rca --html report.html --open  # a single self-contained HTML report
rca --pr 16889 --fail-on high  # exit non-zero when QA risk is high
rca serve --repo /path         # local web UI with a base picker

CI gate: risk via exit code only, no PR comments, no writes. Copy examples/github-actions/culprit-pr.yml into .github/workflows/:

- run: pip install "culprit>=0.3.0"
- env: { GH_TOKEN: "${{ github.token }}" }
  run: rca --pr ${{ github.event.pull_request.number }} --fail-on high

HTML report

--html writes one self-contained file. No CDN, opens offline, attaches to CI. It opens with the verdict: a scored QA risk with the factors behind it, and the prime suspect with how long the bug lived.

Then it reconstructs how the bug got there. Every commit that touched the buggy line, from creation through the commit that broke it (red, with the exact diff) to the fix (green):

Every node expands to its diff. See the full report.

vs git bisect

--bisect "<cmd>" runs a real bisect as an optional confirmation layer, in a throwaway git worktree so your checkout is never touched. When the first failing commit matches the blamed suspect, the report stamps it confirmed by git bisect.

Architecture

One normalized context in, one structured JSON result out. The only non-deterministic step is the optional LLM narrative, isolated behind an adapter so the engine runs with no API key.