Giving AI agents dermatology tools over MCP

An agent that answers questions about skin needs more than a model. It needs tools that return structured, cited knowledge, and it needs those tools to refuse the things it should not do. DermLiving offers both through the Model Context Protocol, an open protocol for connecting AI applications to tools and data.

One endpoint, no install

The tools are served remotely at https://api.dermliving.com/mcp, over MCP's streamable HTTP transport. There is nothing to install and no key to request. Add the endpoint to any client that supports remote MCP servers:

{  "mcpServers": {    "dermliving": { "type": "http", "url": "https://api.dermliving.com/mcp" }  }}

The endpoint is stateless, answers in JSON and supports MCP protocol versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05. Its test suite drives it with the official MCP TypeScript client.

The five tools

Every tool is read-only and declares no side effects: none of them changes anything. Each tool publishes a JSON Schema for its arguments, and arguments that break it are refused before anything runs.

What a call looks like

Under the client, a tool call is one JSON-RPC request. You can send one by hand:

curl https://api.dermliving.com/mcp \  -H 'content-type: application/json' \  -H 'accept: application/json, text/event-stream' \  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "map_terminology", "arguments": {"text": "eczema"}}}'

The result arrives twice: as text for the model to read, and as structuredContent for code. Here is the structured part, shortened to the first candidate:

{  "trace_id": "trc_ZvmoI6N4LC",  "knowledge_version": "kv_2026_10_01",  "result": {    "input": "eczema",    "ambiguous": true,    "note": "Lay umbrella term. Confirm the intended concept.",    "candidates": [      { "concept_id": "cond_atopic_dermatitis", "name": "Atopic dermatitis", "match": "lay_term", "score": 0.81 }    ]  }}

The ambiguous flag is the tool telling the agent to ask before it acts. What does eczema mean? explains why.

The same rules as the REST API

The MCP endpoint is not a second implementation. Each tool call goes through the same gateway as POST /v1/agent-tools/execute, with the caller's own credential:

• Scopes: a tool runs only when the caller holds its scope. Callers without a key get the sandbox's scopes, which cover all five tools.
• Rate limits: the same limits as the REST API, counted per caller.
• Abstentions: ask_knowledge refuses clinical decisions and patient data exactly as POST /v1/knowledge/ask does, with the same reason codes.
• Tracing: every result carries a trace id and the knowledge version it read, so an answer can be traced back to the call that produced it.

Beyond MCP

Agents that prefer plain HTTP can use the REST API, described in OpenAPI 3.1. The API also publishes /.well-known/agents.json at api.dermliving.com, and the website publishes llms.txt for models reading the docs. All of them describe the same ten routes and five tools.

Like the REST API, the MCP endpoint serves the synthetic sandbox today. Its answers show the format and the boundaries, not clinical content.