@pipeworx/mcp-vc-finder

vc-finder

VC / angel investor graph: which investors fit a raise, who is on which round, and which investing person at a firm to approach — every fact carrying the page it came from. Built for Mojibake's own raise first (Bruce ask 2026-09-29) and public since fleet #2663 (Bruce 2026-10-01: "just make it live; it will be easier to test"). Not npm-published — that is a separate Bruce-gated path. See docs/vc-graph-plan.md, docs/vc-graph-architecture.md, docs/vc-finder-staging-contract.md.

Part of Pipeworx — an MCP gateway connecting AI agents to 1866+ live data sources. This is an independent, unofficial integration — not affiliated with, endorsed by, or published by the upstream provider.

Store: D1, not Postgres (fleet #2526 store-change)

The architecture doc (§1) originally called for a dedicated Supabase Postgres project. Superseded 2026-09-30 by the PM (Bruce ruled out new Supabase spend): the store is ONE Cloudflare D1 database. Schema: sql/vc-finder-schema.sqlite.sql — a narrow port (only the tables this pack's 8 tools need) written because Liesel's full port was not yet on origin/main when this was built in parallel. Reconcile the two when hers lands; delete this one rather than let two schema files drift.

Shape

• mcps/vc-finder/ — this pack. 10 tools, vc_ prefixed. Calls a _vcIndex service-binding-shaped object (RPC methods, not fetch — see VcIndexBinding in src/index.ts), the same shape as the POD_INDEX precedent (workers/pod-index, docs/pod-researcher-design.md).
• workers/vc-index/ — the pipeworx-vc-index worker skeleton. A WorkerEntrypoint exposing one RPC method per tool, backed by src/queries.ts (parameterized SQL against the D1 binding — no stored procedures; D1 has none). Deployed (fleet #2579) against the pipeworx-vc-finder D1 database; CI redeploys the code on push (.github/workflows/deploy-vc-index.yml) and never touches the data.
• scripts/vc-finder/fixtures/seed.sqlite.sql — fixed-id fixtures (a16z, Sequoia, SignalFire, two companies, two rounds, claims with provenance, precomputed investor_sector_portfolio_counts / investor_observed_stats rows) for local testing only.

Local test (never a remote D1 database)

wrangler d1 execute VC_INDEX_DB --local --cwd workers/vc-index \
  --file sql/vc-finder-schema.sqlite.sql -y
wrangler d1 execute VC_INDEX_DB --local --cwd workers/vc-index \
  --file scripts/vc-finder/fixtures/seed.sqlite.sql -y

Then either query directly (wrangler d1 execute VC_INDEX_DB --local --cwd workers/vc-index --command "...") or run the pack's callTool() against the resulting local D1 sqlite file (workers/vc-index/.wrangler/state/v3/d1/...) via a harness that mocks _vcIndex with src/queries.ts's functions — see fleet #2526's task record for the exact harness used to produce the acceptance evidence (compiled with tsc, run with Node's node:sqlite).

People and contact routes (fleet #2561)

Bruce's 2026-09-30 case ruling: investing people are served with the contact routes they published. vc_search_people and vc_get_person are the two people tools; vc_get_investor lists a firm's partners with their routes, and vc_find_investors_for_company names a relevant_partner per firm when one is known (most senior investing role, then one with a published route). Every path serves only exposure_hold = 0 people and checks the removal list (removal_entry, mirrored from scripts/vc-finder/removal-list.json) at read time. Tests: workers/vc-index/src/queries.people.test.mts. Load/ER details: docs/vc-finder-staging-contract.md "How the store loads and serves them".

Source and removal, stated where a caller can see it (required now that the pack is public — CLAUDE.md "Personal data", Bruce's case ruling). People and their contact routes come from public sources only: the firm's own site or the person's own published page. Every route carries the page it was published on, the quoted text showing it, and a personal-or-role label. Phone numbers and home addresses are never collected. Anyone can request removal at [e-mail] — add them to scripts/vc-finder/removal-list.json and the read path drops them immediately; it does not wait for the next load. The vc_search_people and vc_get_person descriptions carry this same statement, so it reaches a caller who never reads this file, and tests/scope-is-not-a-gate.test.ts fails if either description loses it.

Public since fleet #2663 — what the private gate was, and what replaced it

Through fleet #2579 this pack was unlisted in the manifest, listed in PRIVATE_PACKS, hidden as not found from everyone outside BOARD_OWNER_CALLERS + internal traffic (NOT_FOUND_PRIVATE_SLUGS in workers/gateway/src/index.ts), and re-checked the same rule inside the pack via assertAllowed(). All four are gone. Any caller can call every tool.

NOT_FOUND_PRIVATE_SLUGS is now an EMPTY set rather than deleted — it is the next private pack's gate, and with no members privatePackVisible() answers true for every slug, so a green test over it proves nothing about the not-found path. That is said out loud at the set's definition.

What protects the people in the graph is therefore no longer the gate; it is the read-time removal check below, which was always the real control.

Refreshing the deployed data

The D1 database holds a SERVING copy, never the full local build:

node scripts/vc-finder/build-deploy-copy.mjs "$VC_DATA/<build>.sqlite" "$VC_DATA/vc_finder_deploy.sqlite"
node scripts/vc-finder/d1-export-chunks.mjs "$VC_DATA/vc_finder_deploy.sqlite" "$VC_DATA/d1-import"
node scripts/vc-finder/d1-import.mjs "$VC_DATA/d1-import"      # resumable; verifies row counts

build-deploy-copy.mjs drops every held person (exposure_hold = 1) and every row that exists only because of one, person-flagged companies, and merged-away companies nothing references, and refuses to finish if any held person or orphan survives. The import targets a fresh database (the schema file is plain CREATEs); to refresh, create a new database, import, then point database_id at it.

Known gaps (v0 skeleton, not fixed by this task)

• vc_find_investors_for_company still scores every VC-shaped organization per call, but in a fixed number of set queries (fleet #2579 — it used to be three D1 queries per candidate, which on the real store is past D1's 1,000-queries-per-invocation limit). A materialized stated-fit candidate table would cut the rows read.
• Money fields are REAL — fine for whole-dollar amounts in these fixtures, worth revisiting (integer cents) once real amounts include cents.
• No loader/insert path here — this is the READ side only, per the task.

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "vc-finder": {
      "url": "https://gateway.pipeworx.io/vc-finder/mcp"
    }
  }
}

What this endpoint actually serves

tools/list at https://gateway.pipeworx.io/vc-finder/mcp returns the tools in the table above plus the shared Pipeworx meta-tools — ask_pipeworx, discover_tools, search_within, remember/recall and the rest of the gateway-wide set. So the tool count you see is larger than this table: a single-pack endpoint currently lists roughly 30 shared tools alongside the pack's own. The connection's initialize response states its exact scope, and is the authoritative answer for a given day.

This is deliberate, not multiplexing by accident. The meta-tools are what let a scoped connection answer a question this pack does not cover — via ask_pipeworx, which routes across the whole catalog — without you adding a second MCP server. There is currently no way to mount a pack endpoint without them; if the extra schemas cost you more context than the routing is worth, connect to the full gateway once rather than to several pack endpoints.

Or connect to the full Pipeworx gateway to get every pack's tools listed directly, instead of just this one's:

{
  "mcpServers": {
    "pipeworx": {
      "url": "https://gateway.pipeworx.io/mcp"
    }
  }
}

Both URLs reach the same gateway and the same 1866+ data sources. The only difference is which pack's tools are listed directly; ask_pipeworx reaches all of them from either one.

No MCP client? Call it over HTTP

curl -X POST https://gateway.pipeworx.io/v1/tools/vc_search_investors \
  -H 'Content-Type: application/json' \
  -d '{"stage":"seed","geo":"US","limit":5}'

No account needed for the first calls. Inspect any tool: GET https://gateway.pipeworx.io/v1/tools/vc_search_investors. Find one: POST https://gateway.pipeworx.io/v1/tools/search_packs with {"query":"..."}.

Run it from npm

@pipeworx/mcp-vc-finder is a local stdio bridge to the Pipeworx gateway: your MCP client starts it with npx, and it forwards every call to the same https://gateway.pipeworx.io/vc-finder/mcp endpoint shown above. It needs network access; there is no offline mode.

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

Without a key it calls anonymously, capped at 50 calls a day. A free account raises that to 200 a day: get a key at https://pipeworx.io/signup and add "env": { "PIPEWORX_API_KEY": "<your key>" } to the server entry. The key is sent as a Bearer token, the same way a hosted connection sends it, so calls are metered and billed exactly as they are there.

It lists only this pack's tools by default. Set PIPEWORX_META_TOOLS=1 to also list ask_pipeworx and the other gateway-wide tools.

Behind an HTTPS proxy: Node's built-in fetch does not read HTTPS_PROXY by default. Newer Node releases honour it when NODE_USE_ENV_PROXY=1 is set; check your Node version's documentation.

Using with ask_pipeworx

Instead of calling tools directly, you can ask questions in plain English — this works on the pack endpoint above as well as on the full gateway:

ask_pipeworx({ question: "your question about Vc Finder data" })

The gateway picks the right tool and fills the arguments automatically.

More

• Docs and guides
• pipeworx.io

License

MIT