Archestack MCP server

Archestack MCP server

Ask your assistant about open-source AI projects and get answers from the Archestack rankings: scores, GitHub facts, reviews and open-source picks for paid tools.

It covers two lists:

• Best of AI: self-hosted AI apps and infrastructure, like chat UIs, model servers, RAG, agents and vector databases.
• Best of Vibe Coding: starters, coding agents, MCP servers, skills and plugins for building with AI.

It reads the open data API. No key, no account. Run it on your machine with npm, or add the hosted server by its URL.

Hosted server

https://mcp.archestack.com/mcp: the same tools over streamable HTTP, for apps that take a URL instead of a command. Read-only, no sign-in.

• Claude (claude.ai and the desktop app): Settings → Connectors → Add custom connector, paste the URL.
• Claude Code: claude mcp add --transport http archestack https://mcp.archestack.com/mcp
• Any client: { "type": "http", "url": "https://mcp.archestack.com/mcp" }

It runs on Cloudflare's free Workers plan, so it allows 120 requests a minute per address and can run out of requests late in a busy day (UTC). It keeps no logs of your questions: privacy. For heavy use, run the npm package below.

Safety allowlist for admins

The hosted server also serves Archestack's safety checks as allowlists: MCP servers whose latest run passed every applicable check, pinned to the package version that was read, each with the official MCP Registry's record of that version.

• GitHub Copilot ("Registry only" policy): MCP Registry URL https://mcp.archestack.com/registry (v0.1 API: /v0.1/servers, /v0.1/servers/{name}/versions/{version}).
• Claude Code: managed-mcp.json for a fixed set, or allowedMcpServers for managed settings.
• The data: /api/v1/allowlist.json. A check result is a fact about what was read on that date, not a guarantee.

Install

Needs Node.js 22 or newer. One click for Cursor or VS Code:

Claude Code

claude mcp add archestack -- npx -y @archestack/mcp

Claude Desktop: add this to claude_desktop_config.json (Settings → Developer → Edit Config), then restart the app.

{
  "mcpServers": {
    "archestack": { "command": "npx", "args": ["-y", "@archestack/mcp"] }
  }
}

Cursor: the same mcpServers block in ~/.cursor/mcp.json, or in .cursor/mcp.json for one project.

VS Code without the button:

code --add-mcp '{"name":"archestack","command":"npx","args":["-y","@archestack/mcp"]}'

From source

git clone https://github.com/archestack/mcp.git archestack-mcp
cd archestack-mcp
npm ci && npm run build

Then point your client at node /full/path/to/archestack-mcp/dist/index.js instead of npx -y @archestack/mcp.

Tools

Tool	What it does
search_projects	Finds projects by words and filters: list, category, license family, runs without a GPU, arm64 build, live demo
get_project	One project in full: score and its parts, place in its category, GitHub facts, health, links, review
deploy_facts	What it takes to run one project: Docker, arm64, GPU, RAM, outside services, port, its stated needs, upkeep
top_in_category	The ranking of one category, best first
compare	2 to 4 projects side by side
check_repo	Before you install a GitHub repo: listed or not, license, upkeep, safety checks passed and failed, higher picks
alternatives_to	Open-source picks for a paid product, such as Cursor or ChatGPT

All seven are read-only. Things to ask:

• "Which vector database runs without a GPU?"
• "Open-source alternatives to Cursor"
• "Compare Open WebUI and LibreChat"
• "Top 5 coding agents with an arm64 build"
• "Check github.com/ruvnet/ruflo before I install it"

What comes back, shortened (search_projects with "vector database" and no_gpu, data of 2026-10-13):

{
  "updated": "2026-10-13",
  "total": 15,
  "results": [
    {
      "name": "Milvus",
      "place": 1,
      "score": 81,
      "stars": 46366,
      "license": "Apache-2.0",
      "gpu": "optional",
      "arm64": true,
      "tagline": "Distributed vector database with dense, sparse and hybrid search at scale",
      "source_url": "https://archestack.com/best-of-ai/projects/milvus/",
      "checked_at": "2026-10-13"
    },
    { "name": "Meilisearch", "place": 2, "score": 75, "gpu": "none", "...": "..." }
  ]
}

How it works

• It talks to archestack.com and nothing else. The API is static JSON served with the site. The server sends nothing about you or your questions and has no telemetry.
• The npm server checks each file against the JSON Schema published next to it (/api/v1/schema/). If a file does not match, you get an error instead of a made-up answer. The hosted server skips this check to stay inside the free plan's CPU time; the tests run both against the same recorded files.
• Each request times out after 8 seconds, with two retries and backoff. The server keeps each file in memory for an hour; Archestack refreshes the data twice a day.
• Every project in an answer carries source_url, its page on archestack.com, and checked_at, the date its facts were checked on GitHub.
• ARCHESTACK_API_URL points the server at another copy of the API. The tests use it for a local one.

Data and license

The code is MIT. The data is CC BY 4.0 by Archestack: if you publish something built on it, link the project page (source_url).

Archestack checks every project on GitHub: stars, last commit, license, Docker files. How the score works: archestack.com/methodology/.

Development

npm ci
npm test           # builds, then runs the server over stdio against recorded API answers
npm run fixtures   # records fresh answers from archestack.com into test/fixtures
npx @modelcontextprotocol/inspector node dist/index.js   # try the tools by hand
npx wrangler@4.141.0 dev   # the hosted server on http://localhost:8787/mcp

src/worker.ts is the Cloudflare Worker entry (config in wrangler.jsonc); npm run deploy publishes it with CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID set.

Bugs and ideas: open an issue.