agentsclimarketplace

Release mcp server

Skill tlennon-ie/neurodock/.claude/skills/release-mcp-server

Walk through cutting a PyPI release of one of the five MCP servers — bump version, write CHANGELOG, run tests and evals, tag, publish.From its SKILL.md

Install
npx -y skills add tlennon-ie/neurodock --skill release-mcp-server

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 20 stars20 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.7 KB, as published. Nobody here has run it

release-mcp-server

A checklist for releasing any of the five Python MCP servers published to PyPI:

  • packages/mcp-chronometric/ to neurodock-mcp-chronometric
  • packages/mcp-cognitive-graph/ to neurodock-mcp-cognitive-graph
  • packages/mcp-guardrail/ to neurodock-mcp-guardrail
  • packages/mcp-task-fractionator/ to neurodock-mcp-task-fractionator
  • packages/mcp-translation/ to neurodock-mcp-translation

Releases are CI-driven. You bump versions in a PR, merge it, then push a single v* tag. The Release workflow does the publishing. The contributor reading this should already have decided the version bump via the version-impact skill.

When to use

  • After a PR has been merged that warrants a release of one or more of the servers above.
  • When version-impact flagged needs-bump against a server and you are picking the release up.

Prerequisites

  • The bump kind decision (major / minor / patch) for each affected server, run version-impact first if unsure.
  • Push access to main and permission to push tags. No PyPI or registry token is needed locally, CI holds those.
  • uv available for running tests and evals before you cut the release.

How releases work

The whole release is triggered by pushing one v* tag (for example v0.7.3). That tag fires the Release workflow in .github/workflows/release.yml, which runs three jobs:

  1. npm publishes the changed TypeScript packages with pnpm -r publish.
  2. pypi builds and publishes every Python package with uv build and uv publish, one package at a time.
  3. mcp-registry runs after pypi and publishes every packages/*/server.json to registry.modelcontextprotocol.io using GitHub OIDC. It first rewrites each server.json version to match that package's pyproject.toml, so the manifest and the package stay in lockstep.

You do not run uv publish or touch the registry by hand. Your job is to land the version bump in a PR and push the tag.

What it does

Walks through the release steps in order. The skill does not execute them, it lists them so nothing is skipped.

1. Bump the version (in a PR)

For each affected server, edit packages/mcp-<server>/pyproject.toml and update the version field. Follow semver, the bump kind was decided by version-impact.

A version bump is required even for a docs-only or README change you want the registry to pick up. The registry verifies PyPI ownership via the mcp-name: marker in the published PyPI README, and that README only goes live when a new version is published. Without a bump, the marker never updates.

2. Update the server manifest

The mcp-registry job syncs packages/mcp-<server>/server.json to the pyproject.toml version automatically at publish time, so you do not have to hand-edit the version number. Do update any other fields in server.json that the release changes (new tool descriptions, changed runtime arguments, and similar).

3. Update the local CHANGELOG

Append a ## [<new-version>] - YYYY-MM-DD heading to packages/mcp-<server>/CHANGELOG.md (Keep a Changelog style, matching the existing entries), then ### Added / Changed / Fixed sub-sections with one-line bullets. Plain, factual, ND-readable.

4. Run tests

uv run --directory packages/mcp-<server> pytest

Must exit 0. Coverage gate per pytest.ini for the server.

5. Run evals (if the server has them)

mcp-translation has evals in packages/evals/corpora/translation/. Any prompt change ships through the eval pipeline (plan.md §7):

uv run --directory packages/evals pytest -k translation

Other servers have no eval corpus today, skip this step.

6. Open the PR and merge it

Open a PR with the version bumps, the server.json edits, and the CHANGELOG entries. Get it reviewed and merged into main. Nothing publishes from the PR itself, the publish happens on the tag.

7. Tag the release and let CI publish

From an up-to-date main:

git tag vX.Y.Z
git push origin vX.Y.Z

A single repo-wide v* tag triggers the Release workflow. The pypi job publishes every package whose version is not yet on PyPI, then the mcp-registry job publishes every server.json. Both jobs skip versions that already exist, so an unchanged package is a clean no-op.

Confirm the run in the Actions tab, then check https://pypi.org/project/neurodock-mcp-<server>/<new-version>/ and the server's listing on the MCP registry.

Post-release

  • Update any pyproject.toml in other packages that pinned the released server.
  • If the release added a tool or changed a schema, regenerate any downstream client bindings.
  • Note the version in the root CHANGELOG.md under the next repo-wide release entry (batched, not per-PR, see version-impact skill).

Limitations

  • Publishing is CI-driven and version-gated. If you forget to bump a version, the pypi and mcp-registry jobs skip that package and nothing ships, no error tells you the bump was missing.
  • The registry publish depends on the PyPI publish, the mcp-registry job needs pypi because ownership is verified via the mcp-name: marker in the published PyPI README (ADR 0008 / 0009).
  • Yanking a bad release requires manual PyPI dashboard action, the skill does not cover rollback.
  • Releases are cut from a v* tag on main. Tagging from a branch is possible but not documented here.

Voice

A release is a procedure. List the steps, say what to verify between them, and stop. Resist the urge to write release-note copy in this checklist, that belongs in the CHANGELOG itself.

Keep looking

Skills are one crate of 610,677. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.