TL;DR. The Python quelltest repository has a local MCP server and a Claude Code plugin, so a coding agent can find untested edge cases and write pytest tests that passed all five gates, in-session and without sending code anywhere but your configured LLM provider. Install the mcp extra, point your MCP client at the server, and ask in plain English.
Updated 2026-10-20. This post describes the Python package. The five-tool server and the quelltest-mcp command below are in the repository (version 2.0.1); the PyPI release, 2.0.0, ships an earlier server with different tools (verify_test, get_survivors, generate_killing_test, get_quell_score), started with python -m quell.mcp_server. We checked this by inspecting the published wheel. If you want the newer server today, install from the repository. For JavaScript and TypeScript, the npm package has its own MCP server: see the npm MCP post.
What does the setup look like?
Starting with the repository version, Quelltest ships a local MCP server inside the pip package. Two steps and your coding agent can operate it directly:
pip install "quelltest[mcp]"
// .mcp.json in your project root
{ "mcpServers": { "quelltest": { "command": "quelltest-mcp" } } }
That's the whole setup. Now ask Claude Code to "find untested edge cases in this project and fix the payment ones" — and it does, in-session, without you touching a terminal.
Why did we change direction?
The first version of our Claude integration was a cloud connector: sync your readiness reports up, and Claude.ai could answer questions about your PRS trend. It worked. It also missed the point.
Reading a score in a chat window makes Claude a spectator. What I actually wanted — what I kept doing manually all day — was Claude as an operator: write a function, immediately check what's untested, write the proven test, move on. The loop belongs inside the coding session, next to the code, not in a dashboard.
MCP made the right version almost embarrassingly simple. The server is a thin stdio wrapper over the same SDK the CLI uses. There is no cloud component. There is no network call. The agent spawns a process, the process reads your working tree, results go back over stdio, done.
Which tools does the agent get?
| Tool | What the agent gets |
|---|---|
find_edge_cases | Documented requirements with no test — with file:line locations |
write_verified_tests | A pytest test per gap, each proven through all 5 gates, written to disk |
get_prs_score | Production Readiness Score with tier and per-file breakdown |
prove_file | Coverage ratio for one file or one function |
reproduce_bug | A natural-language bug description turned into a verified failing test |
The important part isn't the tool list — it's what the tools refuse to do.
write_verified_tests will not write a test that didn't pass Gate 5, where
we inject the violation into your source and require the test to fail. An
agent calling this tool cannot flood your repo with decorative tests. The
pipeline physically won't let it.
That matters more in the agent era, not less. Agents generate tests at 50x human speed, and unverified tests at 50x speed is 50x more false confidence. Verification is the part that has to stay ruthless.
Is there a Claude Code plugin?
If you live in Claude Code, the plugin bundles a skill that teaches Claude the workflows — scan after editing, score before a PR, reproduce bugs as failing tests:
/plugin marketplace add quelltest/quelltest-lib
/plugin install quelltest@quelltest
Then /quelltest find src/, /quelltest score, or plain English.
What happened to the cloud connector?
Still there, still useful for the team-dashboard case — PRS trends across projects, readiness questions in Claude.ai chat. But it's now the optional add-on, not the story. The story runs on your machine.
Full setup docs: quelltest.com/docs/mcp. If you hit a rough edge, open an issue — this is a fresh surface and I want the bug reports.
What are the limits?
- The tools run the Python engine, so they only help on Python projects. For JavaScript and TypeScript, use the npm package.
write_verified_testswrites only tests that passed all five gates. Cases the rule engine cannot handle are scaffolded or flagged, not written.reproduce_bugturns a bug description into a failing test and needs LLM authorization configured, asquell reproducedoes.- The Production Readiness Score is a review aid, not a quality guarantee.
- This is a fresh surface. Expect rough edges, and see the changelog.
FAQ
Does the MCP server send my code anywhere?
No. It is a stdio process that reads your working tree. Quelltest transmits code only to the LLM provider you configure, and only when the rule engine is insufficient.
Do I need a terminal?
No. The agent calls the tools itself. The CLI is still there when you want it.
Which Python version is required?
Python 3.11 or newer.
Where is the full setup?
In /docs/mcp, and the Python basics are in the quickstart and how it works.
Sources
- Quelltest on PyPI: pypi.org/project/quelltest
- Python repository, including the MCP server and plugin: github.com/quelltest/quelltest-lib
- Model Context Protocol: modelcontextprotocol.io
- Claude Code MCP documentation: code.claude.com/docs/en/mcp