Synapse.org MCP server icon

Synapse.org

by Synapse.org

HIPAA CompliantSOC2 ReadyISO 27001 Ready
Healthcare5 tools

Search Synapse research datasets and read entity metadata, annotations and provenance from an AI agent. 5 tools, all read-only in source. OAuth requests the view scope only, which cannot download file contents.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect Synapse.org via MCP

https://mcp.synapse.org/mcp

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

Synapse.org Tools & Capabilities (5)

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • Metadata only — no file contents, and no table rows. The view scope and the downloadFile=False calls both stop at metadata. To read a file you still download it through Synapse's own clients, having met that dataset's conditions for use.
  • Synapse's own documentation never mentions MCP. We searched the complete docs.synapse.org corpus — 415 KB of llms-full.txt, roughly 50,000 words — and found zero occurrences of "MCP" or "Model Context Protocol". Every piece of connector-specific documentation lives in the third-party GitHub repository, not on Sage's documentation site.
  • The repository is a personal account, not the Sage Bionetworks organization. The directory attributes the connector to Sage Bionetworks and links documentation and support at github.com/susheel/synapse-mcp. Sage Bionetworks maintains a Sage-Bionetworks GitHub organization referenced elsewhere in its own docs. We could not establish the governance relationship between the two, and state the discrepancy rather than resolving it.
  • The code has not moved since 2025-10-20. That is the last commit on the default branch as of 2026-08-22 — roughly ten months. The deployed version matches it, so the server is consistent with its source rather than silently ahead of it, but neither has changed.
  • A different project owns the synapse-mcp name on PyPI. That package, version 1.3.0, is a "persistent MCP server — session memory, engineering pipeline" tool from an unrelated author and repository. It is not this connector. Installing it would get you something else entirely.
  • Search returns at most 100 results per call, defaulting to 20, and the source clamps the limit rather than erroring on a larger request. Exhaustive enumeration of a large project is not what this tool does.
  • Search return fields are fixed. The source hardcodes name, description and node type, and silently retries without return fields if Synapse rejects them, surfacing a warning in the response. You cannot request other fields through the tool.
  • get_entity_children works on projects and folders only. Anything else returns a stated error rather than a result.
  • No tool-level permission check exists in the server. A has_scope helper is defined in source and never called by anything, and the auth middleware's own documentation describes 403 responses for insufficient permissions as "handled at tool level" — a layer that is not there. Authorization is entirely Synapse's, applied to your token. This does not widen what you can reach, but the server enforces no second boundary of its own.
  • Four unreachable read methods exist in the entity layer. get_file_content_url, query_table, get_table_columns and get_dataset_items are defined and called by no tool. They are dead code today, and get_file_content_url in particular is a file-download path with no tool wired to it. Nothing exposes them — but they indicate a direction of travel that a future release could reach, and a view-scoped token should not be assumed sufficient for it.
  • The maintainers advise caution with consumer AI services. The repository's compliance notice states Synapse Terms of Service prohibit redistribution, that provider-side storage or logging "could be interpreted as unauthorized redistribution", and marks consumer cloud AI under "use with caution". Its recommended alternatives are enterprise deployments with contractual data protection or self-hosted models. The connector's own authors flag this connector's most common deployment as the risky one, which is unusually candid and should be read before use.
  • We could not read a live tools/list response. The endpoint returned 401 to an anonymous request, so the annotations reported here come from the server's source at the version the deployment reports, not from a handshake.
  • No prompts capability. Neither the directory nor the source registers any prompt, and the source contains no @mcp.prompt decorator.

Frequently asked questions

It can reach the metadata, not the files. Synapse defines the view permission as seeing that an entity exists plus its annotations, explicitly not file contents. The connector requests only view, and its entity reads pass downloadFile=False in source. Controlled-access governance still applies to metadata, which Synapse warns must not be emailed or redistributed.

No. Synapse's authorization server advertises nine scopes including download, modify and authorize, and the MCP server requests two of them: openid and view. That is a real boundary between the routine read and the consequential one, and it is one of the few connectors we have checked where the narrow scope was actually taken.

No. All five tools declare readOnlyHint true and destructiveHint false in the server's source, and no code path calls a Synapse write endpoint. The one POST request is hardcoded to the search endpoint, where POST carries the query rather than a change. The view scope would not authorize a write in any case.

The server's own maintainers say to be careful. Its compliance notice lists consumer cloud AI services under use with caution, because Synapse Terms of Service prohibit redistribution and provider-side storage may count as that. It recommends enterprise deployments with data-retention guarantees or self-hosted models instead.

Five, and this is one of the rare cases where two independent enumerations agree exactly. Anthropic's directory lists five names, the repository README documents the same five, and the source registers exactly five tool decorators. The deployed server reports the same version as the commit we read, so the match is verified rather than assumed.

One resource, which the directory does not list. The source registers synapse://feeds/blog, returning live RSS XML from the Sage Bionetworks publication feed. It fetches a third-party URL and passes the response into agent context unparsed, so treat that feed as untrusted input rather than as Synapse data.

Yes. The endpoint returned HTTP 401 to an anonymous request on 2026-08-22 with a Bearer challenge, so every session authenticates. Sign-in is OAuth through Synapse by default. What you can see is whatever your own Synapse account can see, since the server applies no permission filter of its own.

Anthropic's directory attributes it to Sage Bionetworks, and the linked repository is a personal GitHub account rather than the Sage Bionetworks organization. The code is MIT licensed and its last commit was 2025-10-20. Beware a name collision: the synapse-mcp package on PyPI is an unrelated project by a different author.

Sources

Use in Agentman

Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio

Server Info

Category
Healthcare
Developer
Synapse.org
Tools
5
Domain
mcp.synapse.org

Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.