Autism Research Knowledge Graph (MCP)

An MCP server over a knowledge graph built from 1,216 peer-reviewed research papers on autism and neurodevelopmental research. Connect any MCP client — Claude, or your own agent — and query the literature directly.

Read this before you buy

The corpus is narrower than "autism research" suggests. It centres heavily on motor control, praxis and neuromotor work: imitation, postural control, gait, cerebellar and mirror-neuron research. Measured by entity coverage, motor topics outweigh social-cognitive ones roughly 65 to 1.

If your work is on motor development, praxis, coordination or cerebellar involvement in autism, the depth here is unusual.

Every search returns a relevance score, and a weak best match is flagged in a warning field rather than dressed up as an answer.

Tools

Tool	What it does
search_papers	Hybrid BM25 + semantic search across the graph. Sub-second.
get_entity	Look up a named concept — a brain region, syndrome, assessment instrument, theory. Exact-title match preferred.
get_community_overview	Survey a topic at a chosen granularity, from broad themes to specific clusters.
list_entities	Browse entities by type. Metadata listing, not a relevance search.
deep_search	Synthesise an answer across multiple sources. Several seconds.

Ten entity types: BRAIN_REGION, BODY_STRUCTURE, MOTOR_PHENOMENON, SYNDROME, SYMPTOM_DIMENSION, ASSESSMENT_INSTRUMENT, EXPERIMENTAL_PARADIGM, ANALYSIS_TECHNIQUE, FINDING, THEORY.

What it is built on

Source PDFs were processed within a GraphRAG: entity extraction, community detection, and per-entity summary cards. The result — 258,278 artifacts spanning entity cards, community summaries and source passages — lives in a vector store supporting both keyword and semantic retrieval.

deep_search retrieves on two legs — graph-level summaries for synthesis and the underlying source passages for detail — then synthesises across both.

Source documents are not identified. The corpus was assembled under access that permits querying it, not redistributing it, and some of the underlying documents are copyrighted. Results carry no filenames or identifiers, and deep_search attributes findings generally — "one study found", "several studies report" — rather than naming authors, journals or years. If you need to cite a specific source in your own work, this index will point you at the finding, not at the paper.

Pricing

Event	Price
search — any lookup, browse, or search call	$0.01
deep-search — multi-source synthesis with a language model	$0.05

deep_search is priced higher because it runs a language model over the retrieved context. The other four tools are sub-second retrieval.

Connecting

The Actor runs in Standby mode as an always-ready HTTP server. Point an MCP client at the Standby URL with /mcp as the path, using your Apify API token as a Bearer token.

Known limits

• Coverage is uneven. See the scope note above.
• No source citations. See above — results identify findings, not documents.
• Entity typing is imperfect. Types were assigned by a language model over extracted text and corrected in a later pass, but edge cases remain — some cerebellar cell types sit under BRAIN_REGION where BODY_STRUCTURE would arguably be better.
• This is a research index, not a clinical resource. Nothing it returns is medical advice.