MCP Servers

MCP (Model Context Protocol) servers are external tool servers that expose capabilities to your agents via a standardized protocol. ElasticClaw can start MCP servers as subprocesses inside each agent sandbox and register their tools with the agent's gateway.

Supported sources:

• npx — NPM packages (e.g. @modelcontextprotocol/server-github)
• uvx — Python packages via uv
• smithery — Smithery registry packages
• docker — Docker images
• sse — Server-Sent Events endpoints (remote)

Configure MCP server definitions in the server settings UI under Settings → MCP Servers. Workspaces opt into the servers they need from elasticclaw-config.yaml.

mcp_servers:
  - name: github
    source: npx
    package: "@modelcontextprotocol/server-github"
    enabled: true
    config:
      repository: "elasticclaw/elasticclaw"
    secrets:
      GITHUB_TOKEN: github_mcp_token

  - name: postgres
    source: docker
    image: mcp/postgres
    enabled: true
    config:
      database_url: "postgresql://localhost/mydb"

  - name: remote-tools
    source: sse
    url: https://tools.example.com/sse
    enabled: true

The secrets map under each MCP server maps environment variable names to secret names configured for that MCP server in settings.

mcp_servers:
  - name: github
    secrets:
      GITHUB_TOKEN: github_mcp_token

Add mcps to your workspace'selasticclaw-config.yaml to enable specific MCP servers for agents created from that workspace:

# elasticclaw-config.yaml
mcps:
  - name: github
    config:
      repository: "my-org/my-repo"   # workspace-level override

Each agent will start the configured MCP servers as subprocesses and register their tools with the OpenClaw gateway.