@pharmatools/opengate-mcp

An MCP server that lets an AI agent check whether its own answers are grounded in the context it was given — before returning them.

It exposes one tool, check_grounding, backed by OpenGATE's deterministic grounding check. No LLM-as-judge: the verdict is reproducible and gold-anchored.

Why

An agent doing RAG or document QA can hallucinate a figure, drop a required fact, or answer confidently when the context doesn't actually contain the answer. check_grounding catches all three, deterministically:

• Required facts you name must appear in the answer.
• Every number in the answer must trace to the context (or the question).
• Unanswerable questions must be declined, not fabricated.

Add it to your MCP client

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "opengate": {
      "command": "npx",
      "args": ["-y", "@pharmatools/opengate-mcp"]
    }
  }
}

Then, in a conversation, the agent can self-check:

Before you answer, use check_grounding with your draft answer and the retrieved passages.

The tool

check_grounding

| Argument | Type | Meaning | |---|---|---| | answer | string | The answer to check. | | context | string | string[] | The retrieved context it must be grounded in. | | question | string (optional) | The original question (lets its numbers count as grounded). | | expected_facts | string[] (optional) | Facts a correct answer must contain. | | allowed_new_numbers | string[] (optional) | Numbers the answer may introduce that aren't in the context. | | must_abstain | boolean (optional) | Set true when the context can't answer — the answer must then decline. |

Returns a GROUNDED / NOT GROUNDED verdict with named issues, plus structured output (grounded, issues, anchorsMissed, ungroundedNumbers, abstained).

Privacy

Runs entirely on your machine. check_grounding is a deterministic, local text comparison — the answer and context you pass are processed in memory and never transmitted or stored. No accounts, no analytics, no external network calls, no third-party model. Full policy: https://www.pharmatools.ai/privacy

Related

• OpenGATE (the framework): https://github.com/nickjlamb/opengate
• Evaluate a whole gold set in CI: Getting Started

MIT licensed.