MCP Server Setup

Two ways to connect: Streamable HTTP (zero install, remote) or stdio via uvx (local process). Both support the full 13-tool surface.

Prerequisites

  • A Synapse Layer account at forge.synapselayer.org
  • A Connect Token (sk_connect_*) from Dashboard → Connect
  • For stdio: Python 3.10+ and uvx (comes with uv)

Remote (Streamable HTTP)

No local installation required. Point your MCP client to the hosted endpoint:

json
{
  "mcpServers": {
    "synapse-layer": {
      "url": "https://forge.synapselayer.org/api/mcp",
      "transport": "streamable-http",
      "headers": {
        "x-connect-token": "sk_connect_your_token"
      }
    }
  }
}

Note

Streamable HTTP is the recommended transport for most use cases. It requires no local Python installation and connects directly to the Forge infrastructure.

Local (stdio via uvx)

Run the MCP server locally as a subprocess. Useful for Claude Desktop and Cursor:

json
{
  "mcpServers": {
    "synapse-layer": {
      "command": "uvx",
      "args": ["synapse-layer"],
      "env": {
        "SYNAPSE_API_KEY": "sk_connect_your_api_key",
        "SYNAPSE_AGENT_ID": "my-agent"
      }
    }
  }
}

The uvx command downloads and runs the synapse-layer package from PyPI in an isolated environment. No manual pip install required.

Environment Variables

VariableRequiredDescription
SYNAPSE_API_KEYYes (stdio)Connect token for API authentication (sk_connect_*)
SYNAPSE_AGENT_IDNoAgent identifier. Default: default-agent
LOG_LEVELNoLogging verbosity: DEBUG, INFO, WARNING, ERROR. Default: INFO
SYNAPSE_PRIVACY_EPSILONNoDifferential privacy epsilon. Range: 0.01–10.0. Default: 0.5

Verify Connection

After configuring your MCP client, ask the agent to run health_check. A successful response confirms the connection:

json
{
  "status": "healthy",
  "version": "2.4.x",
  "engine": "synapse-memory",
  "user_id": "your-tenant-id"
}