MCP server
Use MentionBeat from Claude, Claude Code, or any MCP client — read your measurement data
and work the action queue without leaving your editor.
Install — build it once, then point your MCP client at the file
(
claude_desktop_config.json, or
.mcp.json for Claude Code):
# in the repo
cd packages/mcp-server && npm install && npm run build
{
"mcpServers": {
"mentionbeat": {
"command": "node",
"args": ["/absolute/path/to/packages/mcp-server/dist/index.js"],
"env": {
"MEASURE_API_URL": "http://localhost:8000",
"LLMVIS_API_KEY": "<your org API key>",
"STUDIO_URL": "http://localhost:4444",
"MENTIONBEAT_TOKEN": "<your platform token>"
}
}
}
}
Everything runs under your own key and token — the server scopes every call to exactly what you can already
see and do. MENTIONBEAT_TOKEN is only needed for the action tools.
The package isn't on npm yet, so this is a local path for now.
Tools
| Tool | What it does | Writes? |
|---|
list_projects | Every brand/product being tracked. | — |
get_project | One project: prompt counts, runs, latest run id. | — |
list_runs | Measurement runs, newest first, with actual cost. | — |
get_run_metrics | Every metric with its 95% CI and sample size. | — |
get_citations | Domains the engines cited, classified own/rival/third-party. | — |
get_benchmark | You vs the field: mention share and rank. | — |
get_hallucinations | Engine claims that contradict your approved facts. | — |
get_trends | Headline metrics over time, with noise-aware deltas. | — |
get_experiments | Measured lift from a specific content change. | — |
list_actions | The work queue: what to fix, ranked, with evidence. | — |
get_fix | Paste-ready JSON-LD / FAQ / intro for one page. | — |
check_ai_access | Whether AI crawlers can actually reach a site. | — |
run_site_audit | Crawl + audit a site, adding findings to the queue. | yes |
mark_action_fixed | Mark an item fixed — triggers an immediate re-check. | yes |
verify_actions | Re-check everything marked fixed. | yes |
The loop this makes possible. Ask your assistant to "fix my top three visibility
findings": it calls list_actions, pulls the exact markup with
get_fix, applies it in your repo, then calls
mark_action_fixed — which re-audits the live page and reports back whether the
finding is genuinely gone. A fix that didn't land comes back as not cleared, with the reason, and
the item returns to the queue. The assistant never gets to mark its own homework.