# MCP Starter Pack

Bootstrap Claude Desktop or Cursor with a minimal, useful MCP set — then attach our model bridges.

## What you get

- `mcp.json` — example server block you can merge into Claude Desktop (`claude_desktop_config.json`) or Cursor MCP settings
- Scoped filesystem, fetch, and memory
- Placeholders for **cheaperllm** and **anymodl** HTTP MCP endpoints

## Setup (Cursor)

1. Install Cursor with our referral: https://cursor.com/referral?code=BVN3MVJH9SDO
2. Copy the `mcpServers` object from `mcp.json` into your Cursor MCP config
3. Set env vars:
   - `CHEAPERLLM_API_KEY` — from https://cheaperllm.com/?utm_source=agentstackkit&utm_medium=site&utm_campaign=hub
   - `ANYMODL_API_KEY` — from https://anymodl.com/?utm_source=agentstackkit&utm_medium=site&utm_campaign=hub
4. Restart the editor and run `mcp-probe` (see /tools/) against each server

## Setup (Claude Desktop)

1. Open Claude Desktop settings → Developer → Edit Config
2. Merge `mcpServers` from `mcp.json`
3. For Claude **API** usage in agents (not the Desktop UI), prefer cheaperllm Claude routes: https://cheaperllm.com/?utm_source=agentstackkit&utm_medium=site&utm_campaign=claude-api
4. Product UI (no affiliate): https://claude.ai/?utm_source=agentstackkit&utm_medium=referral&utm_campaign=stack

## Credits for chat

Need a multi-model chat while wiring tools? OneAIPass: https://oneaipass.com/?utm_source=agentstackkit&utm_medium=site&utm_campaign=hub

## Security notes

- Never point filesystem MCP at `$HOME` — use a project root only
- Do not commit API keys; use environment variables
- Prefer read-only tools until the agent proves it can follow tool discipline (`/packs/prompt-ops/02-tool-discipline.md`)

## Next

- Full MCP catalog: https://agentstackkit.com/mcp/
- Agent loop pack: https://agentstackkit.com/packs/workflow-agent-loop/
- Machine catalog: https://agentstackkit.com/api/catalog.json