crowllm-mcp
Official# crowllm-mcp
MCP server for [CrowLLM](https://crowllm.com), an OpenAI-compatible gateway with one key for 180+ models and a zero-rated `:free` shelf.
## Tools
- `search_models` — search the live catalog (`GET /v1/models`), optionally free ids only
- `get_pricing` — list prices in USD per 1M tokens from `GET /api/pricing`
- `chat` — send a prompt to any model through `POST /v1/chat/completions` and get the reply with usage
## Install
```bash
npm install -g crowllm-mcp
```
or run without installing:
```bash
npx -y crowllm-mcp
```
or from a clone:
```bash
git clone https://github.com/crowwdev/crowllm-mcp.git
cd crowllm-mcp
npm install
node index.js
```
## Configure (Cline / Claude Desktop / any MCP client)
```json
{
"mcpServers": {
"crowllm": {
"command": "npx",
"args": ["-y", "crowllm-mcp"],
"env": {
"CROWLLM_API_KEY": "your-crowllm-key"
}
}
}
}
```
- `CROWLLM_API_KEY` — created at [crowllm.com](https://crowllm.com); required for `chat`, optional for `search_models` and `get_pricing`
- `CROWLLM_BASE_URL` — optional, defaults to `https://crowllm.com`
## Notes
- Use `search_models` with `free_only: true` to see ids that bill at zero.
- Forked/free ids are rate-limited; use them for exploration rather than long loops.
## License
MIT
## Registry
This repository carries a `server.json` manifest for the official MCP registry
(`io.github.crowwdev/crowllm-mcp`, npm package `crowllm-mcp`, stdio transport). It becomes
publishable once the npm package is released.
TDQS
Scored across 3 tools
Each tool maps to a distinct stage of the workflow: discovering models (search_models), comparing cost (get_pricing), and actually running inference (chat). Search and pricing both touch the model catalog, but the descriptions make the split clear (ids vs. USD per 1M tokens). No plausible misselection.
search_models and get_pricing follow a clean verb_noun snake_case pattern, and chat is a single word that fits naturally. Minor deviation only: chat lacks a noun object, but it is unambiguous in context.
Three tools is on the lean side but proportionate to a thin LLM gateway surface: discover, price, invoke. Nothing feels padded, though the set is small enough that a little more breadth would be welcome.
The core loop of finding a model, checking its price, and chatting is fully covered. Gaps are minor and workaroundable: no embeddings/completions-only endpoint, no account balance or usage-history lookup, and no way to fetch a single model's full metadata beyond what search returns.