AI Investing Tools That Work with Claude, Cursor, or MCP

AI tools, Claude Desktop, Claude Code, Cursor, Windsurf, are excellent for investment-memo writing and synthesis but blind to current data. Their training cutoffs are months old; they cannot see this week's GitHub commits, this week's startup signals, or this week's funding patterns. The Model Context Protocol (MCP) solves this by letting you plug structured data tools into the AI's runtime.

For VC research the canonical install is the GitDealFlow MCP server (@gitdealflow/mcp-signal on npm). Six read-only tools cover the full research workflow:

- `get_trending_startups`, top 20 startups across all sectors by GitHub commit-velocity acceleration - `search_startups_by_sector`, filter by one of 15 sector slugs (ai-ml, fintech, devtools, security, data-infra, etc.) - `get_startup_signal`, look up an individual startup's current metrics by name - `get_signals_summary`, dataset freshness and counts - `get_scout_receipts`, grade a GitHub user's starring history against a curated database of validated unicorns - `get_methodology`, full methodology document

Install on Claude Desktop: add {"mcpServers": {"gitdealflow": {"command": "npx", "args": ["-y", "@gitdealflow/mcp-signal"]}}} to claude_desktop_config.json. Restart Claude Desktop. The six tools appear automatically.

Install on Claude Code or Cursor: the same MCP server works, see the install instructions in the package README.

Install on any MCP host: the server works with any client that speaks Model Context Protocol over stdio.

HTTP fallback for non-MCP hosts: POST https://signals.gitdealflow.com/api/mcp/rpc exposes the same toolset over Streamable HTTP transport. Useful for OpenAI Assistants API, Gemini function-calling, custom orchestration.

Workflow examples. Ask Claude "which AI/ML startups are accelerating most this week?" and Claude calls get_trending_startups filtered to ai-ml. Ask Claude "what is $org's engineering acceleration vs the sector?" and Claude calls get_startup_signal plus search_startups_by_sector to compare. Ask Claude "is this founder a real scout?" with a GitHub username and Claude calls get_scout_receipts to grade their starring history.

The tools are free in perpetuity. There is no API key gate. The free tier is structurally permanent.

Read-only design matters more than it first appears. Because the six tools cannot write, mutate, or trigger any action, an AI agent holding the server can only query, never change, your data. There is no API key to leak, no billing credential to store, and no rate-limit gate that silently degrades a long research session. The server runs locally on your machine, so requests originate from your own environment rather than a third-party relay.

Privacy behaves the same way. The server makes outbound calls to the GitDealFlow public dataset endpoint, and the endpoint logs only anonymous request metadata such as the path, the status code, and response time, for debugging purposes. It does not log the calling host or any user identifier, and the MCP server itself ships no telemetry. For a fund doing quiet sourcing work, that is a meaningful difference from hosted analytics tools that keep a record of every query.

The package carries two independent trust signals worth noting. It is published in the official Model Context Protocol Registry under io.github.kindrat86/vc-deal-flow-signal, and it holds an A-tier rating on Glama, the leading MCP server directory. Both listings describe the same six read-only tools and the same no-key install, which makes the tool easy to audit before wiring it into a production research workflow.

For hosts that do not speak MCP natively, the HTTP fallback is the bridge. ChatGPT does not support MCP at the consumer tier, but the same toolset is exposed over Streamable HTTP transport at the /api/mcp/rpc endpoint, callable from custom GPT actions or the Assistants API. The same route serves Gemini function calling and custom orchestration, so a single research question can be answered through Claude or Cursor via MCP and through ChatGPT via HTTP with identical underlying data.

Composition is where the value compounds. A typical morning is one query for the top accelerating names in a sector, a second query to pull a specific startup's signal against its sector, and a third to grade a founder's scout history. Because each tool returns structured data rather than prose, the results can feed directly into a memo, a tracking sheet, or a scoring pipeline without a manual copy-paste step.