Hlido MCP Server is an independent AI agent trust scoring, claim auditing, and comparison tool that can be queried by other agents via the MCP (JSON-RPC) protocol. Hlido is positioned as an independent AI agent evaluation platform, similar to a "Rotten Tomatoes for AI agents," publishing evidence-backed scorecards based on manual testing, including 0–100 scores, grade determinations, item-by-item claim PASS/FAIL audits, and signed screenshots. This repository contains the source code of the MCP server that exposes this registry to other agents.
In terms of purpose, this server is used to let AI agents programmatically query Hlido's public evaluation data, including trust scores, claim audit results, and comparison information between agents.
Main capabilities include: providing independent trust scores, claim audits, and comparison queries; access via the MCP protocol using JSON-RPC; and also including discovery/metadata tools, with the specific tool table available in src/index.mjs.
In terms of usage, the hosted server can be used directly without installation. The server runs as a Cloudflare Worker at hlido.eu/mcp and is suitable for Claude Desktop, Cursor, or any MCP client (configured via mcpServers). The service is also listed on Smithery and mcp.so. For local running, Docker can be used; the container runs the worker on the local workerd runtime via wrangler dev, without requiring a Cloudflare account, and the MCP endpoint is http://localhost:8080/（GET to get server information, POST to make JSON-RPC calls). For self-deployment, this is a standard Cloudflare Worker; copy wrangler.toml.example to wrangler.toml, set the account id, then run npx wrangler deploy. Optional bindings (KV cache, Vectorize similarity index) degrade gracefully when missing, and the worker falls back to directly fetching public JSON.
In terms of design principles: only public data is used; the server reads the same JSON as hlido.eu/data/* (registry, scorecards, proofs), never exposes scoring weights, grader assertions, or editorial drafts; the methodology remains private, while results and evidence are public; no authentication is required and no tracking is done, using anonymous JSON-RPC, with only lightweight daily per-tool counts retained as telemetry; it is deliberately kept lightweight—this is an adapter for open data, while the evaluation pipeline, testing engine, and scoring model are located elsewhere and are not part of this repository.
In terms of data and licensing: the code uses the Apache-2.0 license; the evaluation data uses the CC-BY license via the public data mirror and HF dataset. Related links include the website https://hlido.eu、智能体清单 https://hlido.eu/agent-manifest.json、llms.txt https://hlido.eu/llms.txt、公开数据镜像 https://github.com/ankitkapur1992-hlido/hlido-public。
Applicable scenarios include: AI agents querying other agents' trust scores and audit results during operation; developers or platforms integrating Hlido's public evaluation data; accessing trust scores and comparison information through MCP clients (such as Claude Desktop, Cursor).
In terms of limitations: the server only provides public data and does not expose scoring weights, grader assertions, or editorial drafts; the evaluation pipeline, testing engine, and scoring model are not in this repository; the optional KV cache and Vectorize similarity index degrade to directly fetching public JSON when missing; the specific tool list must be checked in src/index.mjs.
