wafergraph-mcp

# wafergraph-mcp

A remote MCP server exposing [wafergraph.com](https://wafergraph.com)'s semiconductor & AI
supply-chain dataset — 565 companies across 12 segments, the supplier/customer graph, and a
74-deal M&A corpus — as nine read-only tools any MCP-speaking AI agent can call directly.

No auth, no cost, read-only. Streamable HTTP transport at `/mcp`. Human landing page at `/`.

Live: **https://wafergraph-mcp.jwpalm99.workers.dev**

## Tools

| Tool | Purpose |
|---|---|
| `search_companies({query?, segment?, country?})` | Name/description search with segment & country filters. Compact list, capped at 25, with a total match count. |
| `get_company({id})` | Full allowed profile for one company (by id or name) plus its supplier/customer edges. |
| `get_segments()` | The 12-segment taxonomy (+ subsegments, market_position enum) with live company counts. |
| `get_supply_chain({id, direction, depth})` | Walk the supplier/customer graph from a focal company up to 2 tiers up (`"up"`), down (`"down"`), or `"both"`. |
| `get_deals({query?, segment?})` | Search the M&A deal corpus. Compact list, capped at 30, with a total match count. |
| `compare_companies({ids})` | Side-by-side comparison of 2-6 companies on aligned fields, plus their shared and unique supply-chain counterparties. |
| `get_country_exposure({segment?})` | Geographic concentration: which countries host a segment's companies, ranked by count, with disclosed market-cap coverage. |
| `find_chokepoints({segment?, limit?})` | Rank chokepoints by downstream dependency weighted by market position. Transparent heuristic, not a proprietary risk model. |
| `analyze_portfolio_exposure({holdings})` | Map tickers or ids to segment/country exposure and flag suppliers that several holdings share. Not investment advice. |

Every response includes:
- `data` — the payload.
- `attribution` — compiled-by/sources block (SEC, Wikidata, Wikipedia, GLEIF) plus a
  `company_url` back to the full sourced profile on wafergraph.com.
- `links` — `{ report, newsletter }` pointing back to wafergraph.com and its $99 report.

### Field discipline

Company records include `key_products` as of 2026-07-19. That field in wafergraph's own dataset
was originally bulk-drafted with a known fabrication/mis-scope history and was withheld here
until a verified deep-fill landed upstream; it now has (confirmed live) 564/565 companies filled,
with only `sk_enpulse` empty (defunct, absorbed into SKC 2025-12-23). Every field shipped here —
`name`, `ticker`, `market_cap_usd_b`, `segments`, `one_liner`, `market_position`, `key_products`,
`key_customers`/`key_suppliers` graph edges, deals — is established, trust-checked data. Fields
are still whitelisted (not blacklisted) in `src/types.ts` (`toAllowedCompany`), so anything added
upstream later stays excluded by default until deliberately added.

## Install

### Claude Code

```
claude mcp add --transport http wafergraph https://wafergraph-mcp.jwpalm99.workers.dev/mcp
```

### claude.ai (custom connector)

Settings → Connectors → Add custom connector → paste:

```
https://wafergraph-mcp.jwpalm99.workers.dev/mcp
```

### Generic MCP client (Streamable HTTP)

```json
{
  "mcpServers": {
    "wafergraph": {
      "url": "https://wafergraph-mcp.jwpalm99.workers.dev/mcp"
    }
  }
}
```

## Architecture

Cloudflare Worker, TypeScript, built on the `agents` package's `McpAgent` (Durable-Object-backed
Streamable HTTP MCP server — free on the Workers Free plan; SQLite storage backend). See
`CLAUDE.md` for the file map and the data-source-mode decision (hybrid live-fetch + one
vendored snapshot, discovered by E2E-testing the upstream endpoints rather than assumed).

## Development

```bash
npm install
npm run dev       # wrangler dev, local
npm run typecheck # tsc --noEmit
npm run deploy    # wrangler deploy
```

## Attribution & licensing

See `ATTRIBUTION.md` for the full source/license breakdown (SEC EDGAR public domain, Wikidata
CC0, Wikipedia CC BY-SA 4.0 — attribution required, GLEIF). Short version: free to use, credit
wafergraph.com, not financial advice.

This is an independent project built against wafergraph.com's public dataset; it is not an
official wafergraph product.