> ## Documentation Index
> Fetch the complete documentation index at: https://docs.llm-stats.com/llms.txt
> Use this file to discover all available pages before exploring further.
# MCP Server
> Query the LLM Stats catalog and call the gateway from any MCP-compatible client.
The LLM Stats MCP server gives AI agents access to the model catalog, benchmark scores, rankings, and pricing over the [Model Context Protocol](https://modelcontextprotocol.io/). One workflow tool (`compare-run`) also runs prompts through the inference gateway and returns latency, tokens, and cost per model.
The server is hosted and read-only over the [Stats API](/api-reference/introduction). Cursor, Claude Code, Claude Desktop, Continue, and ChatGPT Apps connect directly.
## Setup
Point your MCP client at the hosted server. No installation required.
Add this to your Cursor MCP settings (`.cursor/mcp.json`):
```json theme={"theme":{"light":"github-light","dark":"github-dark"}}
{
"mcpServers": {
"llm-stats": {
"url": "https://mcp.llm-stats.com/mcp",
"headers": {
"Authorization": "Bearer ze_"
}
}
}
}
```
```bash theme={"theme":{"light":"github-light","dark":"github-dark"}}
claude mcp add llm-stats --transport http https://mcp.llm-stats.com/mcp \
--header "Authorization: Bearer ze_"
```
Add this to `claude_desktop_config.json`:
```json theme={"theme":{"light":"github-light","dark":"github-dark"}}
{
"mcpServers": {
"llm-stats": {
"command": "npx",
"args": ["-y", "mcp-use", "client", "connect", "https://mcp.llm-stats.com/mcp"],
"env": { "MCP_AUTHORIZATION": "Bearer ze_" }
}
}
}
```
Add this to your Continue `config.json`:
```json theme={"theme":{"light":"github-light","dark":"github-dark"}}
{
"experimental": {
"modelContextProtocolServers": [
{
"name": "llm-stats",
"url": "https://mcp.llm-stats.com/mcp",
"headers": { "Authorization": "Bearer ze_" }
}
]
}
}
```
Configure your app with the hosted MCP URL and store the Bearer token in your platform's secrets manager. ChatGPT renders every widget the server exposes.
```
URL: https://mcp.llm-stats.com/mcp
Header: Authorization: Bearer ze_
```
Any client that supports HTTP transport works. Set the server URL to `https://mcp.llm-stats.com/mcp` and pass your API key in the `Authorization: Bearer ` header.
## Authentication
The MCP server uses the same `ze_...` key as the [Stats API](/api-reference/introduction#authentication) and the [Gateway API](/gateway/introduction#authentication). You don't need to provision a second key. Get yours from the [developer console](https://llm-stats.com/developer).
The same `ze_...` Bearer token authenticates the Stats API, the inference
gateway, and the MCP server. Provision once, use everywhere.
## Tools
### Read tools
Read tools are safe to call at any time. They don't modify state and don't consume gateway credits.
| Tool | Description |
| ----------------- | ---------------------------------------------------------------------------------- |
| `list-models` | List models with filters (organization, modality, price, context, sort, cursor). |
| `get-model` | Full detail for one model id: every benchmark score, provider pricing, and source. |
| `list-benchmarks` | Benchmark catalog. Use it to disambiguate benchmark ids. |
| `list-scores` | Score matrix with filters across models and benchmarks. |
| `get-rankings` | TrueSkill rankings for a category. |
| `list-updates` | Models added in the last N days (1–30). |
### Workflow tools
Workflow tools chain multiple API calls. They're still read-only.
| Tool | Description |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------- |
| `compare-models` | Side-by-side comparison of 2–4 model ids. |
| `find-best-models-for-category` | Top-ranked models for a category, optionally filtered by price or open-weight. |
| `resolve-model` | Fuzzy lookup: free text in, ranked candidate `model_id`s out. Use before any other tool when the id is fuzzy. |
### Gateway tools
Gateway tools call the [inference gateway](/gateway/introduction). The same `ze_...` key authenticates them.
`compare-run` hits the gateway and consumes credits. It's annotated
`readOnlyHint: false` and `openWorldHint: true` so MCP clients can prompt
for approval before calling.
| Tool | Description |
| --------------------- | ----------------------------------------------------------------------------------------------------------------------------------- |
| `compare-run` | Run a single prompt through 2–4 models in parallel; returns per-model latency, tokens, cost, and output. Capped at 2000 max tokens. |
| `get-gateway-snippet` | Return a ready-to-paste curl, Python, AsyncOpenAI, or JavaScript snippet for invoking a model through the gateway. |
## Resources
Seven MCP resources are available for introspection and reference lookups.
| URI | Description |
| ----------------------------- | -------------------------------------------------------------------------------------------------------------- |
| `config://server-context` | Redacted server config: auth mode, base URL, and connection status. |
| `docs://capabilities` | Canonical tool, resource, prompt, and widget inventory with annotations. |
| `data://benchmarks` | Full benchmark catalog (`{ id, name, category, modality, max_score, verified }`) so the LLM picks correct ids. |
| `data://categories` | Category ids (`coding`, `math`, `reasoning`, …) usable with `get-rankings`. |
| `data://organizations` | Organization catalog (`{ id, name }`) derived from the model index. |
| `data://providers` | Provider catalog (`{ id, name }`) derived from every model's `providers[]`. |
| `recipe://gateway-quickstart` | Paste-ready gateway quickstart: base URL, auth, OpenAI compatibility, streaming, tools, multimodal, snippets. |
## Slash-command prompts
Four MCP prompts show up as slash-commands in clients like Cursor and Claude Desktop.
| Prompt | Arguments | What it does |
| ---------------------- | ---------------------------------- | -------------------------------------------------------------------------------------------------------------------------------- |
| `/compare-models` | `model_a`, `model_b` | Resolves each name with `resolve-model`, then calls `compare-models` with the resolved ids and writes a structured comparison. |
| `/find-cheapest-for` | `task`, optional `max_input_price` | Maps the task to a category via `data://categories`, calls `find-best-models-for-category`, explains the price/quality tradeoff. |
| `/whats-new-this-week` | (none) | Calls `list-updates` with `days=7` (falling back to 14), summarizes notable releases grouped by org. |
| `/explain-benchmark` | `benchmark_id_or_name` | Disambiguates via `data://benchmarks`, fetches top scores via `list-scores`, writes a SOTA and saturation summary. |
## Interactive widgets
Seven tools render an interactive widget alongside a markdown summary. MCP Inspector, ChatGPT Apps, Cursor, Claude Desktop, VS Code, and Goose render both. Text-only clients fall back to the markdown.
| Widget | Backing tool(s) |
| ----------------------- | ---------------------------------------------- |
| `leaderboard` | `list-models`, `find-best-models-for-category` |
| `model-card` | `get-model` |
| `model-compare-table` | `compare-models` |
| `benchmark-leaderboard` | `list-scores` (sorted, single-benchmark) |
| `rankings` | `get-rankings` |
| `updates-timeline` | `list-updates` |
| `compare-run` | `compare-run` |