devstack-mcp

A remote MCP (Model Context Protocol) connector for DevStack API — the libraries.io successor: one keyless, normalized view of the whole developer registry across npm, PyPI, Docker Hub, and the VS Code Marketplace (search, package details, versions, download/pull stats), plus resolved dependency graphs (deps.dev), per-version vulnerabilities (OSV.dev), OSSF Scorecard insights (deps.dev), and cross-registry lookups over 50+ ecosystems (ecosyste.ms).

Target deploy: https://devstack-mcp.vercel.app/mcp — 10 tools. Since the upstream DevStack deployment is metered (RapidAPI/Apify) and gates /v1/* behind a RapidAPI proxy-secret guard, this connector authenticates its own outbound calls with that same proxy secret (DEVSTACK_MCP_PROXY_SECRET, sent as the X-RapidAPI-Proxy-Secret header) and applies a soft per-IP rate limit (DEVSTACK_MCP_RATE_LIMIT, default 30 tool-calls/hour, in-memory) so the free MCP tier stays a discovery channel rather than an unmetered bypass of the paid listing — see lib/ratelimit.js.

What this is, and why it's a separate connector

DevStack API is a plain REST API. Any HTTP client can already call it directly. This repo exists because MCP clients (Claude, ChatGPT, and other MCP-aware agents) don't consume arbitrary REST APIs — they consume MCP tools. devstack-mcp is a thin adapter layer that:

• Exposes each DevStack endpoint as a discoverable, typed MCP tool (name, description, zod input schema, annotations) that an LLM can reason about and call directly, instead of having to be taught the REST surface out-of-band.

• Speaks the MCP streamable-HTTP transport at a single /mcp endpoint, so it can be registered as a connector in Claude, ChatGPT, or any other MCP client with one URL.

• Does nothing else. It has no business logic of its own — every tool call is a pass-through fetch to DevStack, and the JSON response DevStack returns is handed back verbatim as the tool result.

Related MCP server: Developer Tools MCP Server

Authentication: None on the MCP side (deliberate)

DevStack's data has no per-user dimension at all — it's public developer-registry metadata (package listings, versions, download counts, dependency graphs, published CVEs). There is nothing to gate per-caller, so this connector intentionally ships with:

• No OAuth, no login, no per-user bearer tokens

• No Supabase / database

• No demo-vs-real account split — every caller gets the same real, live data

The one server-side secret it carries (DEVSTACK_MCP_PROXY_SECRET) is not per-caller auth — it's how this connector, as a single consumer, gets past the upstream's RapidAPI proxy-secret guard to reach real data. Combined with the soft per-IP rate limit, that keeps this a free discovery tier over the metered DevStack product rather than an unmetered bypass of the paid RapidAPI/Apify listing.

Tool list

One tool per DevStack /v1 endpoint (from devstack-api/openapi.yaml):

All 10 tools are read-only and annotated { readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: true }. scan_vulnerabilities_batch is an HTTP POST but is a read-only query (it mutates nothing), so it carries the same read-only annotations.

How it wraps devstack-api

Each tool handler does a plain fetch(\${DEVSTACK_MCP_API_BASE_URL}${path}`, ...)` against the real DevStack REST API and returns the parsed JSON as MCP tool-result content:

{ content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] }

Upstream HTTP errors (4xx/5xx) are caught and surfaced as a typed MCP error result via an asError helper rather than crashing the request.

Environment variables

Project layout

api/mcp.js       MCP endpoint (StreamableHTTPServerTransport, stateless, no per-caller auth)
api/health.js    GET /api/health
lib/tools.js     All 10 tool definitions (zod schemas + fetch-and-forward handlers)
lib/ratelimit.js Soft per-IP tools/call cap (free discovery tier)
local-server.js  Plain-Node http server for local dev / smoke testing (not deployed)
test/smoke.mjs   Real end-to-end smoke test (initialize, tools/list, tools/call)
vercel.json      Routes /mcp -> api/mcp.js, /health -> api/health.js
server.json      MCP registry publish manifest

Local development

npm install
npm run dev          # starts local-server.js on http://localhost:3900

Endpoints locally: POST http://localhost:3900/mcp, GET http://localhost:3900/health.

To point at a self-hosted / open DevStack instance instead of production (useful for a fully keyless end-to-end run, since production's /v1/* is behind the RapidAPI proxy-secret guard):

# In devstack-api (guard is inert when RAPIDAPI_PROXY_SECRET is unset):
node server.js                                   # listens on :8080

# In devstack-mcp:
DEVSTACK_MCP_API_BASE_URL=http://localhost:8080 npm run dev

Smoke test

npm run dev &                 # terminal 1
npm run smoke                 # terminal 2

test/smoke.mjs drives the running server over real HTTP/JSON-RPC and verifies:

1. GET /health returns { ok: true, ... }

2. initialize succeeds and reports serverInfo.name === "devstack"

3. tools/list returns all 10 tools with their zod-derived JSON schemas

4. tools/call for get_package with { registry: "npm", name: "react" } returns real package data fetched live (not mocked)

Against production, step 4 needs DEVSTACK_MCP_PROXY_SECRET set (the RapidAPI proxy guard); against a locally self-hosted devstack-api with no RAPIDAPI_PROXY_SECRET, the guard is inert and real data flows keyless.

Deploy

Not deployed by this build. Once ready:

cd /Users/isaiahdupree/Software/devstack-mcp
npx vercel --yes --prod

After deploy, set DEVSTACK_MCP_PROXY_SECRET in the Vercel project so the connector can reach the guarded production /v1/*, then register the MCP connector URL in Claude/ChatGPT/any MCP client:

https://<deployment-domain>/mcp