---
title: MCP
description: Connect Claude Code, Claude Desktop, Cursor, or any MCP client to LLMentions over streamable HTTP, and the full tool catalog.
group: Interfaces
---

# MCP

LLMentions is a remote MCP server using the streamable HTTP transport:

- **URL:** `https://api.llmentions.com/mcp`
- **Auth:** `Authorization: Bearer sk_live_…` or `sk_test_…`. [Sign in](https://api.llmentions.com/login) to get your key; see the [Quickstart](/docs/).
- **Server name:** `llmentions`

## Clients

### Claude Code

```bash
claude mcp add --transport http llmentions https://api.llmentions.com/mcp \
  --header "Authorization: Bearer $LLMENTIONS_API_KEY"
```

### Cursor, Windsurf and other JSON-configured clients

Add this to the client's MCP config (for Cursor, `~/.cursor/mcp.json`):

```json
{
  "mcpServers": {
    "llmentions": {
      "url": "https://api.llmentions.com/mcp",
      "headers": { "Authorization": "Bearer sk_live_…" }
    }
  }
}
```

### Clients that only speak stdio

Use the `mcp-remote` bridge:

```json
{
  "mcpServers": {
    "llmentions": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://api.llmentions.com/mcp",
               "--header", "Authorization: Bearer ${LLMENTIONS_API_KEY}"],
      "env": { "LLMENTIONS_API_KEY": "sk_live_…" }
    }
  }
}
```

## Tools

Every tool returns `{data, charged_usd, next}`:
- `data` is concise by default. Pass `response_format: "detailed"` where it is supported to get everything.
- `next` tells the agent what to do next.
- Errors carry a hint, for example: `unknown engine(s): bard — call engines_list for valid keys`.

| Tool | What it does | Cost |
|---|---|---|
| `account_get` | Workspace, key mode (live/test), balance, configured engines, webhook secret. **Call first.** | free |
| `engines_list` | Engine keys and whether each is configured | free |
| `ai_answers` | Ask engines now and get normalized answers and an optional mention report. Synchronous, ≤ 50 checks, nothing stored | per answer |
| `project_list` / `project_create` | Brands you track: name, domain, competitors, prompts | free |
| `prompts_track` | Add or remove tracked prompts | free |
| `visibility_run` | Async panel: prompts × engines × samples. Supports `dry_run`, `queued` and `webhook_url` | per answer |
| `job_get` | Poll a run; returns the report when it finishes | free |
| `visibility_report` | Mention rate with 95% CI, share of voice, per engine and per prompt, top sources, and a significance-tested change vs the previous run | free |
| `answers_search` | The raw stored answers. `missing_target=true` shows where you were absent | free |
| `answers_analyze` | How engines *portray* each brand: sentiment, recommended vs listed, factual claims. Rates with 95% CIs, plus a verifier-agreement quality report | $0.002 / answer |
| `prompts_suggest` | Researcher + critic agents find the buyer questions worth tracking in your category. `track: true` adds them | $0.02 / call |
| `site_audit` | Can AI read your site? Crawls up to 100 pages as AI bots see them (no JavaScript), renders a few in a real browser to find content that needs JavaScript, checks robots.txt per AI bot and whether your CDN blocks them, and lists issues most-severe first with how to fix each. Stored, with what changed since last time | $0.0002 / page, $0.002 / rendered page |
| `audit_get` | A stored audit (by id, or a domain's latest) and what changed since the one before. Poll this for background audits | free |
| `bots_report` | Which search and AI bots really visit, from your server logs: hits, errors and top pages per bot, live AI fetches, and which hits are fakes (IP outside the vendor's published ranges) | free |
| `stats_mention_rate` | Turn *k of n* into a rate with a 95% CI | free |

## A typical agent loop

This is the loop a well-behaved agent runs. Every step is a tool call; nothing needs a human.

1. `account_get`: check the mode and balance.
2. `project_create` with 5–20 buyer-intent prompts.
3. `visibility_run` with `dry_run: true`, then run it for real.
4. Poll `job_get`, or pass a `webhook_url`.
5. `visibility_report`. Act only where `change_significant` is true.
6. `answers_search` with `missing_target: true`, to read the answers that left you out and see what they cite instead.
7. `answers_analyze`, to see whether you are recommended or merely listed, and to check what the engines claim about you.

Start with `prompts_suggest` if you don't know which prompts to track.
