Deepstack MCP Server

DeepStack MCP server

The Model Context Protocol server behind DeepStack, an independent publication on the money behind AI. For each argument about AI capital (capex, credit, power, chips, the labs), DeepStack publishes the strongest case on each side with dated, sourced claims, takes a side (the DeepStack Call), and writes down the test that will settle it: the event, the date, the rule, and the probability the Call gave "yes" before the date. When the date comes, the result is published, whichever way it went.

This server lets any MCP client read that record.

• Endpoint: https://mcp.deepstack.ltd/mcp (Streamable HTTP, JSON-RPC 2.0)
• Authentication: none. Read-only. Free.
• Registry: io.github.mcauduro0/deepstack in the official MCP Registry
• Docs: deepstack.ltd/mcp

Every answer is the published analysis as text, cites the canonical deepstack.ltd URL of each item it names, and ends with "Analysis and opinion, not investment advice." No model runs at request time.

Tools

All tools carry readOnlyHint: true.

Connect a client

Configuration files for each client are in clients/.

Claude (claude.ai and the desktop app). Settings, Connectors, Add custom connector. Name it DeepStack and use the URL https://mcp.deepstack.ltd/mcp, with no OAuth.

Claude Code

claude mcp add --transport http deepstack https://mcp.deepstack.ltd/mcp

Cursor, Windsurf and other editors (mcp.json)

{ "mcpServers": { "deepstack": { "url": "https://mcp.deepstack.ltd/mcp" } } }

VS Code (.vscode/mcp.json)

{ "servers": { "deepstack": { "type": "http", "url": "https://mcp.deepstack.ltd/mcp" } } }

Any client, by hand

curl -s https://mcp.deepstack.ltd/mcp -H 'content-type: application/json' -H 'accept: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"deepstack_docket","arguments":{}}}'

Limits and what is kept

• 60 requests per minute per session (the Mcp-Session-Id that initialize issues), else per address; 30 initialize calls per minute per address.
• The questions asked through the tools are kept as text, with the date, the tool, the number of results and the name the calling program gives itself, to learn which disputes readers ask about. No address, user agent or cookie is kept. Details: deepstack.ltd/privacy.
• Terms: deepstack.ltd/terms.

How it works

A Cloudflare Worker in plain JavaScript (ES modules, no bundler, no npm dependencies). It reads DeepStack's public static API at https://deepstack.ltd/api/v1/ with a ten-minute cache and answers from it. The same JSON is open to any program; this server makes it answerable in a conversation.

src/index.js    router, CORS, cache, rate limit, the questions store
src/mcp.js      JSON-RPC (initialize, ping, tools/list, tools/call) and the eight tools
src/lib.js      pure helpers: keyword index and ranking, excerpts, formatting
src/stats.js    aggregates for the owner's admin route
src/events.js   per-day counts of the site's named controls
schema.sql      the D1 tables
server.json     the MCP Registry manifest

Run the tests with Node 20 or later:

npm test

To run your own copy you need a Cloudflare account: create a D1 database, put its id in wrangler.toml, apply schema.sql, set the ADMIN_TOKEN secret, and npx wrangler deploy. Point API_BASE at any host that serves the same JSON shape.

This repository mirrors the server that runs at mcp.deepstack.ltd. The publication itself, its data pipeline and its editorial process are not part of it.

License

The code is MIT-licensed (see LICENSE). The content the server returns belongs to DeepStack and is governed by the terms.

Analysis and opinion, not investment advice.