Helicone MCP Server

What it does

Helicone is an observability platform that records the calls an application makes to large language models: prompts, responses, cost, latency, errors and custom properties. This server, published by Helicone from the helicone-mcp folder of its main repository, gives an assistant read access to that record through your Helicone API key, so you can ask it to find failing requests, slow or expensive calls, or the steps of one conversation without opening the dashboard.

It has three tools, not the two its README lists. query_requests searches logged requests with Helicone's filter syntax (model, provider, status, cost, latency, user, custom properties, AND/OR combinations), sorts and pages them, returning 100 rows by default and fetching request and response bodies only when includeBodies is set. query_sessions finds sessions within a required time window. The third, use_ai_gateway, is registered in src/index.ts but missing from the README: it posts an OpenAI-format chat completion to ai-gateway.helicone.ai, so the assistant can call models from OpenAI, Anthropic, Google and others through Helicone and have the call logged with an optional session id, user id and custom properties.

The server runs locally over stdio and talks to Helicone's hosted API; it is aimed at teams already sending their model traffic through Helicone.

What it needs

Node.js with npx. A Helicone account and an API key from Settings, API Keys, set as HELICONE_API_KEY; the server exits at start without it. The use_ai_gateway tool also needs the account to be able to pay for model calls through the AI Gateway (credits or provider keys set up in Helicone).

Limits

Request queries go to https://api.helicone.ai and gateway calls to https://ai-gateway.helicone.ai; both addresses are fixed in src/lib/helicone-client.ts with no setting to point at the EU region or a self-hosted Helicone, although the README sends EU users to an EU key page. Results are 100 per call unless limit is changed. The gateway tool reads the reply as a single JSON body, so its stream option is not useful. Version 0.1.6 is an early release.

Who it suits

Developers whose applications already log LLM traffic to Helicone's US cloud and who want to debug errors, cost or latency from their editor. It is mostly read-only, but use_ai_gateway spends money on model calls, so leave it unused if you only want log access. Teams on Helicone's EU region or self-hosting it should note the fixed US addresses. If you do not use Helicone, this server has nothing to read.