pr-docs-mcp

Turn a GitHub pull request into a clear Mintlify documentation draft — using an MCP server for the data and a Claude skill for the writing.

This is an AI-native documentation pipeline: the server fetches the facts, Claude writes the prose, and Mintlify publishes the page. No copy-pasting PR descriptions by hand.

How it works

GitHub PR ──> fetch_pr (MCP tool) ──> doc-drafter skill (Claude writes) ──> docs-site/*.mdx ──> Mintlify
   data            plumbing                    the brain                       the page         the website

• The server is plumbing. docs-pipeline.ts exposes one data-only MCP tool, fetch_pr. It makes no LLM calls — it just returns facts about a pull request.
• The skill is the brain. doc-drafter/SKILL.md tells Claude to call fetch_pr, then write the documentation itself and save it into the docs site.
• Mintlify is the destination. The generated .mdx lives in docs-site/ and renders as a clean docs website.

The fetch_pr tool

Input:

field	type	description
owner	string	repository owner (user or org)
repo	string	repository name
pull_number	positive integer	the PR number

Returns the PR’s number, title, body, changed files, and commits (commit messages) — as JSON.

Input is validated with Zod; failures come back as an MCP isError result rather than crashing the server.

Use it from npm (no clone needed)

The fastest way to use fetch_pr in any project — install straight from npm, no source code required.

1. Go to the project where you want it:

cd /path/to/your-project

2. Register the published server (token pulled from your gh login):

claude mcp add docs-pipeline -e GITHUB_TOKEN=$(gh auth token) -- npx -y pr-docs-mcp

npx -y pr-docs-mcp fetches the published package and runs it — no source code lives in your project.

3. Confirm it’s registered:

claude mcp list

You should see docs-pipeline in the list.

4. Start a fresh Claude Code session in that folder. MCP servers load when a session starts, so reload the VSCode window (or run claude in the folder).

5. Check it’s connected, then use it. Type /mcp — you should see docs-pipeline with the fetch_pr tool. Then ask:

Use fetch_pr to get PR 42 from owner/repo

$(gh auth token) needs the GitHub CLI installed and logged in (gh auth login). No gh? Swap in -e GITHUB_TOKEN=your_token instead.

Optional — the doc-drafter skill makes the assistant write the full doc, not just fetch data. It ships inside the package:

npm install pr-docs-mcp
cp -r node_modules/pr-docs-mcp/doc-drafter ~/.claude/skills/

Then ask: “write docs for PR 42 in owner/repo.”

Other MCP clients (Claude Desktop / Cursor)

Add the published server to the client’s MCP config:

{
  "mcpServers": {
    "docs-pipeline": {
      "command": "npx",
      "args": ["-y", "pr-docs-mcp"],
      "env": { "GITHUB_TOKEN": "your_token" }
    }
  }
}

Restart the client, then ask: “fetch PR 42 from owner/repo.”

Run from source (for development)

Want to hack on the server itself? Clone it instead of installing from npm:

git clone https://github.com/nihaanth/pr-docs-mcp.git
cd pr-docs-mcp
npm install

Add a GitHub token (yours stays local — .env is gitignored):

cp .env.example .env
# edit .env and set GITHUB_TOKEN=...

A token with no scopes is enough for public repos (and raises your rate limit from 60 to 5000 requests/hour). Private repos need the repo scope (classic) or “Pull requests: Read” (fine-grained).

Register the local server — note tsx and the full path instead of npx -y pr-docs-mcp:

claude mcp add docs-pipeline -e GITHUB_TOKEN=your_token -- npx tsx /full/path/to/pr-docs-mcp/docs-pipeline.ts

Local development

npm test           # run the unit tests (Node's built-in test runner)
npm run typecheck  # type-check with tsc
npm run dev        # run the server directly (stdio)

Inspect the tool manually in a browser (no AI client needed):

npx @modelcontextprotocol/inspector npx tsx docs-pipeline.ts

Preview the docs site:

cd docs-site && npx mintlify dev

Releasing to npm

This repo keeps two READMEs: README.md (this one — the full story, shown on GitHub) and npm-readme.md (install-focused, shown on the npm page). npm always publishes the root README.md, so scripts/publish.mjs swaps npm-readme.md into place just for the publish, then restores README.md (the restore runs in a finally, so it happens even if publish fails).

npm version patch            # 1.0.0 -> 1.0.1 (commit + tag)
npm run release              # build, swap in the npm README, publish, restore
npm run release -- --dry-run # same, but `npm pack` instead of publish — inspect before shipping

Project structure

docs-pipeline.ts        MCP server — registers the fetch_pr tool
github.ts               fetchPr() logic + Zod input schema (pure, testable)
github.test.ts          unit tests for fetchPr
doc-drafter/SKILL.md    the Claude skill that orchestrates fetch -> write -> save
docs-site/              Mintlify docs site (docs.json + .mdx pages)
scripts/publish.mjs     publish helper — swaps in npm-readme.md for the npm page
npm-readme.md           install-focused README shown on npmjs.com
.env.example            template for the required GITHUB_TOKEN

Status & roadmap

This is a working portfolio project, not a hosted service. Known limitations and next steps:

• fetch_pr returns at most 30 files / 30 commits (GitHub’s default page size) — pagination (per_page / octokit.paginate) is a planned improvement.
• An evaluation harness for the skill’s output lives under evals/ and is a work in progress.
• docs-site/ ships with just the Mintlify config; pages are generated into it by the pipeline.

Tech stack

TypeScript · Model Context Protocol SDK · Octokit · Zod · Mintlify · Node’s built-in test runner