strand-mcp

Hosted MCP server for Lattice spatial proteomics from H&E. Lattice predicts spatially resolved protein-marker maps from routine H&E, delivered as virtual mIF channels. MCP clients with direct file-transfer network access can browse, upload, estimate or submit, monitor, and download over an OAuth-authenticated Streamable HTTP connection.

📚 Full documentation: https://docs.strandai.com/sdks/mcp

Connect

Add a custom connector pointing at the hosted server URL https://mcp.strandai.com/mcp (transport: Streamable HTTP), then authorize with your Strand AI account and pick an organization. Step-by-step: https://docs.strandai.com/connectors/claude.

There is nothing to install and no API key to manage. Upload and download bytes bypass the conversational model context and MCP transport: upload_sample immediately creates the owned sample and returns a presigned, resumable transfer the client PUTs bytes to directly. Lattice still processes the stored slide to generate predictions. The returned upload_id is the sample id used by get_sample and submit_job; finalizing the transfer starts ingest. Download tools return short-lived signed links.

Agents can browse owned and public samples through one catalog, inspect capped sample job history, read one exact job with get_job, update sample metadata, and estimate a job with submit_job(dry_run=true) before reserving credits. get_job is one-shot: it does not poll, schedule another check, or wake the conversation after a nonterminal response.

Research use only (RUO). Not for use in clinical or diagnostic procedures.

For programmatic or bulk workflows, use the Python SDK or REST API instead.

Analytics and privacy

The hosted server emits one mcp_tool_called operational event per tool invocation. Its closed schema contains only the tool name, a static intent enum, outcome, bounded latency, a coarse client enum, deterministic hashed organization/user/OAuth-client pseudonyms, and a bounded failure category. PostHog person/group profiles and GeoIP enrichment are disabled for these events.

The server never sends tool arguments or results, prompts, filenames, sample or job IDs, marker names, slide data, credentials, OAuth tokens, signed URLs, raw User-Agent values, exception names, or exception messages. Adding an analytics property requires a code and privacy review; unrestricted argument capture is forbidden.

MCP events use Strand's shared PostHog project. That project is currently configured for 84 months of event retention; its API reports events_retention_enforced=false, and there is no shorter MCP-specific TTL. This long shared retention is why the payload is minimized and identifiers are one-way pseudonymized before ingestion. The project owner must reverify and document any retention-setting change. Analytics are used only for product reliability and aggregate connector usage; access is limited to Strand staff with PostHog project access.

PostHog project tokens are public ingestion identifiers, so these events are not a security audit log and can be forged by someone who obtains the browser project token. Security, billing, and authorization decisions must use Strand's server-side records instead. Analytics delivery is best-effort and adds at most 1.25 seconds to a tool response.

Support

Email support@strandai.com.

License

Apache 2.0