MCP setup

Live server: https://syntology.ai/mcp (Streamable HTTP / SSE as your client requires). OAuth discovery is built in — no Syntology-specific client code.

Cursor / Windsurf-style config

{
  "mcpServers": {
    "syntology": {
      "type": "http",
      "url": "https://syntology.ai/mcp"
    }
  }
}

On first use the client runs the OAuth / trial flow. Subscribers get their tier limits; anonymous trial gets the free ceiling.

Trial token without a full OAuth UI

curl -X POST https://syntology.ai/api/oauth/trial/token \
  -H "Content-Type: application/json" \
  -d '{"grant_type":"client_credentials"}'

Hold the token you get (mint rate limits apply per IP).

Which tool for which job

Job	Tool
Free existence check	`have`
Natural-language question	`graph_ask`
Paper by arXiv id	`get_paper`
Declared repos for a paper	`get_code_for_paper`
Repos for a method name	`get_code_for_method`
Verified implementation	`get_reference_implementation` (+ `min_level`)
Browse verified index	`list_reference_implementations`
Citation path	`get_citation_path`
Fit verified routines together	`compose`
Paper methods as typed bricks	`assembly`
Exploratory multi-hop (candidates)	`explore_paths`

Full JSON schemas are published unauthenticated on the MCP server card (generated from the live tool list).

Suggested agent loop

have on the method or arXiv id.
If verified / declared, get_reference_implementation or get_code_for_paper.
Use compose only after you know the output contract you need.
Treat explore_paths as candidates to confirm, not as answers.