devflow-mcp

An MCP server for automating developer workflows across GitHub, Jira, and Slack — built for PR/code review automation (inspect PRs, check CI status, post reviews, sync the linked ticket, notify the team).

Local stdio server: runs on your machine, launched by your MCP client. Not distributed — built for personal/team use.

Tools

Related MCP server: Cursor Auto-Review MCP Server

Automated workflow: PR review

The individual tools above chain into a full PR review workflow — inspect the PR and its CI checks, decide a verdict, post the review, sync the linked Jira ticket, and notify Slack. It's available two ways, at two different layers:

• MCP prompt (pr_review) — defined in src/prompts.ts, part of the MCP protocol itself. It's user-triggered: surfaced as a slash command / menu item by any MCP host that supports prompts (e.g. /mcp__devflow-mcp__pr_review in Claude Code). Works in any MCP client this server connects to, not just Claude Code.

• Claude Code skill (pr-review) — .claude/skills/pr-review/SKILL.md, checked into this repo. It's matched automatically from natural language ("review PR #3 and update Jira") without needing the exact slash command name. Claude Code only.

Both drive the same underlying tools and follow the same 5 steps; the skill is just a Claude-Code-native shortcut for triggering the same workflow implicitly.

Setup

npm install
npm run build

Copy .env.example to .env and fill in credentials, or set these env vars directly in your MCP client config:

• GITHUB_TOKEN — fine-grained personal access token (Settings → Developer settings → Fine-grained personal access tokens), scoped to only the repo(s) this server touches, with permissions: Pull requests (Read and write), Checks (Read-only). Avoid classic repo-scope tokens — they grant full access to every private repo you can see and don't force an expiration.

• JIRA_BASE_URL — e.g. https://your-domain.atlassian.net

• JIRA_EMAIL — the account email tied to the API token

• JIRA_API_TOKEN — id.atlassian.com → Security → API tokens

• SLACK_BOT_TOKEN — api.slack.com/apps → your app → OAuth & Permissions → Bot User OAuth Token (needs the chat:write scope; add chat:write.public too if it should post to public channels it hasn't been invited to)

A tool only fails at call time if its required env vars are missing — you don't need all three services configured to use the others.

Connect to Claude Code

claude mcp add devflow-mcp \
  --env GITHUB_TOKEN=ghp_xxx \
  --env JIRA_BASE_URL=https://your-domain.atlassian.net \
  --env JIRA_EMAIL=you@example.com \
  --env JIRA_API_TOKEN=xxx \
  --env SLACK_BOT_TOKEN=xoxb-xxx \
  -- node /home/channen2/mcp-server/devflow-mcp/dist/server.js

Connect to Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "devflow-mcp": {
      "command": "node",
      "args": ["/home/channen2/mcp-server/devflow-mcp/dist/server.js"],
      "env": {
        "GITHUB_TOKEN": "ghp_xxx",
        "JIRA_BASE_URL": "https://your-domain.atlassian.net",
        "JIRA_EMAIL": "you@example.com",
        "JIRA_API_TOKEN": "xxx",
        "SLACK_BOT_TOKEN": "xoxb-xxx"
      }
    }
  }
}

Develop

npm run dev         # run directly with tsx, no build step
npm run inspector    # open MCP Inspector against the dev server

Capabilities UI

A minimal static dashboard listing every tool, prompt, and skill this server exposes — introspected live from the server (not hand-maintained) plus .claude/skills/*/SKILL.md.

npm run ui   # builds ui/catalog.json, then serves ui/ at http://localhost:4173

Run npm run ui:build alone to just regenerate ui/catalog.json (e.g. after adding a tool).

Next steps

• GitLab wasn't wired up in v1 — same shape (src/<service>.ts + tools in src/tools.ts) if you need it later.

• Distributing beyond your own machine: this stdio server expects Node + env vars on the user's machine. If you ever need to hand this to teammates without a Node setup, repackage it as an MCPB bundle (bundles the Node runtime, installs with a double-click) — ask to use the build-mcpb skill when you're ready.

• Wider/public distribution: if this ever needs to serve multiple users without local installs, move to a remote HTTP deployment (e.g. Cloudflare Workers) — ask to use the build-mcp-server skill again for that path.