Set up the Refactron MCP server for yourself in this repository and prove it
works before you tell me it is done. Do the steps in order.
You are running in Claude Code.
STEP 1 - Check the prerequisites.
Run `node --version`. Refactron needs Node.js 18 or newer. If it is older or
missing, stop and tell me.
Run `python3 --version`, then check that `pytest` and `coverage` are
importable. Coverage attestation is Python-only, and it is what makes a
coverage-backed SAFE verdict possible. Without it the server still runs and
still gates the change, but every verdict caps at UNPROVEN.
STEP 2 - Install the server.
Run `which refactron-mcp`, or `where refactron-mcp` on Windows. If it prints a
path, go to step 3. Otherwise install it:
npm install -g refactron@0.3.0
That puts two binaries on PATH: `refactron` (the CLI) and `refactron-mcp` (the
MCP server). If you cannot install globally, install nothing and use the npx
form in step 3 instead.
Do not use `pip install refactron` for this. The PyPI wrapper provides the
`refactron` command only. It does not ship `refactron-mcp`.
STEP 3 - Register the server.
Run:
claude mcp add refactron -- refactron-mcp
The `--` separator is required. Use `--scope project` instead if I asked for a
config the whole team gets; that writes `.mcp.json` in the repository root.
Without a global install, use this instead. The `-p` flag is required, because
`refactron-mcp` is a second binary of the `refactron` package and not a package
of its own:
claude mcp add refactron -- npx -y -p refactron@0.3.0 refactron-mcp
STEP 4 - Reload.
Claude Code picks the server up on the next session. If you are mid-session and
it does not appear in step 5, tell me to restart `claude`.
A project-scoped `.mcp.json` also needs a one-time approval. If step 5 shows
"Pending approval", tell me to start `claude` in this directory and approve the
server once.
STEP 5 - Confirm the tool is exposed.
Run:
claude mcp list
`refactron` should show as Connected. `claude mcp get refactron` prints the
full record, including scope and resolved command. Inside a session, `/mcp`
lists connected servers and their tools.
You are looking for a server named `refactron`, version 0.3.0, exposing exactly
ONE tool: `verify_change`.
If the server is connected but no tool appears, the handshake failed. Prove the
binary itself works by sending it one initialize request:
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"probe","version":"1.0"}}}' | refactron-mcp
A working server replies on one line with serverInfo.name "refactron". If that
works and the client still shows nothing, the fault is the config: check the
root key and check that the file parses.
STEP 6 - Prove it works with one real verification.
Do not skip this. A connected server is not a working server.
Pick one small source file that this project's test suite exercises. Propose a
real but behaviour-preserving edit to it: reorder the operands of a commutative
expression, rename a local variable, that kind of thing. One or two lines.
Do NOT write that edit to disk. Call verify_change with the proposal instead:
repoRoot the ABSOLUTE path to this repository
edits [{ "path": "<repo-relative path>",
"newContent": "<the full proposed file contents>" }]
testCmd the project's test command, if it is not obvious. For a Python
project installed into the environment (an editable
`pip install -e .` included), prefix it with PYTHONPATH=. so the
tests import the copy being verified rather than the installed
one. Use PYTHONPATH=src for a src layout.
Verify the change BEFORE it exists on disk. The shadow tree is a copy of the
working tree, so if you write the change first and then pass a `git diff` of
it, the diff no longer applies and you get "diff did not apply (stale base?)"
back instead of a verdict.
Verification runs the whole test suite in a shadow copy, so it can take minutes
on a real project. That is the work, not a hang.
STEP 7 - Report, then stop.
Tell me:
- which file you wrote, or which command you ran
- the exact `verdict` and the exact `reason` from the response
- the value of `coverage.tool`: "coverage.py" means coverage was measured,
"none" means it was not
Apply nothing. The step 6 edit was a probe. Confirm with `git status` that the
working tree is exactly as dirty as you found it, and no more.
Do not describe a SAFE verdict as "correct", "proven", or "guaranteed". SAFE
means the project's own tests ran the changed code and stayed green. It
inherits exactly what those tests check, and a weak suite yields a weak SAFE.
MCP