MCP server

The Neurosnap MCP server connects tool-capable clients to account-scoped research operations. It exposes service inspection, supported job submissions, result access, and project operations through Streamable HTTP.

Connect a client

Configure a remote MCP connection to:

https://neurosnap.ai/mcp

Use OAuth when supported by your client, or configure one API-key credential using the client’s custom-header support. See Authentication. A generic REST request to this URL is not a substitute for the MCP initialization and tool-calling protocol.

After connecting, list the available tools. Confirm the account with a non-mutating discovery or history request before authorizing computation. Client menus and JSON configuration formats vary, so follow your client’s remote-MCP instructions.

Discover and validate a service

1. Identify the exact Neurosnap service for the task.
2. Call get_service_spec with its exact title. The result includes the current form and submission_tool name.
3. Use get_service_guidance for tool-specific setup and scientific interpretation.
4. Construct named scalar and file inputs according to the exposed schema.
5. Call validate_job_submission before an unfamiliar submission.

Preflight validates field names and payload shapes without creating a job. File contents receive final validation on submission; passing preflight is not proof that the computation will run successfully.

Submit deliberately

Call the service-specific tool named by submission_tool; do not invent a universal submit tool name. State the authorized candidate count, target, input files, and budget before allowing an agent to execute it.

Use the documented file payload representation. Do not assume a local filesystem path is visible to a remote MCP server. The schema determines whether content, encoded data, or an authorized file reference is accepted.

Record every returned job ID. If a call is interrupted or an error leaves submission uncertain, inspect recent jobs before retrying. duplicate_job intentionally creates another billable run; it does not retrieve a previous result.

Retrieve evidence

Use list_recent_jobs, get_job_details, and get_job_data to recover existing research. File-access tools exposed by the connected server provide supported ways to retrieve artifacts. The read_neurosnap_resource tool can read accessible resource URIs for jobs, projects, or credits.

The available tool list and schemas govern exact arguments. Keep model scores and file provenance with any interpretation. An assistant should distinguish a proposed run from a submitted job and a submitted job from a completed result.

Review access and privacy

The server authenticates requests for the connected account. A client can expose powerful operations, including project changes and billable computation. Review client permissions and what it may transmit to model providers before connecting confidential research.

See Agent data handling and the MCP tutorial. Neurosnap’s built-in Agent and an external MCP client have different execution environments; do not assume their file or network controls are identical.