# Accelerator Application Triage MCP Server (Demo)

A small MCP server that turns startup/founder application triage into
something any MCP-compatible AI client (like Claude Desktop) can query in
plain language, instead of manually digging through a spreadsheet or form
tool.

This was originally built with an accelerator in the food-systems and
plant-based innovation space in mind — the kind of program that runs a
structured pipeline of applications, scoring, and mentor review each cycle.
The tools and mock data reflect that context, but the underlying pattern is
generic: **any accelerator, incubator, or grant program with an application
pipeline could point this same server at their own system of record.**

This is a **portfolio / demo build**, so the mock data here answers questions
specific to that food-systems accelerator context (sector tags like
"alt-protein" and "fermentation," food-sector countries and founders). Swap
`src/data.ts` for a different dataset and the same four tools would answer
the equivalent questions for a fintech accelerator, a climate-tech program,
or any other domain — what changes is the data, not the architecture. In a
real deployment, `src/data.ts` would be replaced with calls to whatever
system of record the organization actually uses (Airtable, Notion, a form
tool's API, etc.) — the four tools below wouldn't need to change at all.
That's the point of MCP: the AI client doesn't care what's behind the tool.

## What it can answer

- "Which alt-protein applications scored 7+ on impact?"
- "What's still sitting in David's review queue?"
- "What hasn't been picked up by any reviewer yet?"
- "Pull up the full file on K4G-2026-058, including notes."
- "Flag Root & Policy Lab for a full panel discussion — team needs
  co-founder diligence."

## Tools

| Tool | What it does | Modifies data? |
|---|---|---|
| `k4g_search_applications` | Filter by text, sector, stage, or minimum impact score | No |
| `k4g_get_application` | Full detail + reviewer notes for one application | No |
| `k4g_list_pending_review` | Everything still needing a decision, optionally by reviewer | No |
| `k4g_flag_for_review` | Marks an application for panel discussion and logs why | Yes |

## Setup

```bash
npm install
npm run build
```

## Running it standalone

```bash
npm start
```

It will sit and wait for MCP messages on stdin/stdout — that's normal, it's
designed to be launched by an MCP client, not run interactively.

## Connecting to Claude Desktop

1. Open Claude Desktop's config file:
   - macOS: `~/Library/Application Support/Claude/claude_desktop_config.json`
   - Windows: `%APPDATA%\Claude\claude_desktop_config.json`

2. Add this server (replace the path with wherever you put this folder):

```json
{
  "mcpServers": {
    "k4g-applications": {
      "command": "node",
      "args": ["/absolute/path/to/k4g-applications-mcp-server/dist/index.js"]
    }
  }
}
```

3. Restart Claude Desktop. You should see "k4g-applications" listed under
   the tools/connectors icon, and you can ask questions like the ones above
   directly in chat.

## Extending this toward a real deployment

- Swap `src/data.ts` for a real API client (Airtable/Notion/etc.) — the
  tool signatures in `src/index.ts` stay the same.
- Add write tools carefully: `k4g_flag_for_review` is deliberately the only
  tool that changes data, and it only ever flags — it never rejects or
  accepts an application outright, so a human still makes the final call.
- For a hosted/shared version (multiple reviewers, not just local use),
  swap the stdio transport for Streamable HTTP — see the MCP docs at
  https://modelcontextprotocol.io.