Read-only MCP server for AI agents

[watilde]

Problem

People already use AI agents to draft phenotypes, but outside WebAPI: the agent guesses concept ids, the JSON is pasted into ATLAS, and nothing records what it chose or why. WebAPI already has the vocabulary, Circe and permissions needed to check and record agent work.

Questions

1. Should an MCP server live in WebAPI, or as a separate service on top of the REST API?
2. If in WebAPI, is a read-only first step OK, targeting 3.1 or later?

First step

Read-only tools, off by default: concept search/details/hierarchy, reading concept sets and cohorts, validating a cohort (Circe + SqlRender, no execution), and reading generation status. /mcp uses the existing auth and @PreAuthorize rules.

Write tools would be a separate discussion, with guardrails: a required rationale and evidence for each write, concepts checked against the vocabulary by id and name, and an append-only audit log.

Cost: one new dependency (Spring AI) and one endpoint. I'm offering to maintain it.

Prototype: #2580 and OHDSI/Atlas3#482. These are references, not merge requests. I'll split or close them based on this discussion.