Skip to content
Ankole

MCP server reference

For AI Agents: the Markdown version of this page is at https://ankole.agentbull.com/en-US/docs/mcp/index.md. The documentation index is at https://ankole.agentbull.com/en-US/llms.txt.

Skills choose domain tools; MCPorter handles discovery and calls. Skill dependencies are not registered as native model tools. Agent Workers call the control plane’s MCP gateway, which resolves credentials and invokes Connector or remote MCP services.

Declare a dependency

Use dependencies.tools in agents/openai.yaml:

dependencies:
  tools:
    - type: mcp
      value: github
      transport: connector
      toolkit: github
      enabled_tools:
        - repos_get
    - type: mcp
      value: lookup
      transport: streamable_http
      url: https://mcp.example.com/mcp
      bearer_token_env_var: LOOKUP_TOKEN
      enabled_tools:
        - search

connector requires toolkit and optionally selects connection and account. Configure tenant applications and personal grants on the Pod Connections page. streamable_http requires url; its optional bearer_token_env_var names a configured WorkerEnv value that the control plane resolves for the proxy. The referenced remote secret is excluded from the Worker’s environment snapshot. Optional protocol_version accepts auto, legacy, or 2026-07-28.

Trusted local servers may use transport: stdio and command. They execute inside Agent Computer through /bin/sh -lc. All transports can declare exact enabled_tools or disabled_tools names. A Skill supports up to 64 dependencies; conflicting declarations for the same server fail setup.

Runtime behavior

Main Agents use enabled any and main Skills. Background Agent Jobs use enabled any and background_job Skills. Each execution receives a disposable MCPORTER_CONFIG with imports: []. Connector and remote entries point to internal gateway endpoints and use a short-lived Turn token, renewed by the Worker. Upstream OAuth tokens and remote bearer values stay in the control plane.

Applet Actions use AnkoleRuntime.mcp directly in Function Runtime; they do not receive an Agent’s Skill environment. Team Actions use the corresponding admin.mcp capability. The trusted host supplies invocation authority outside Applet code.

Call a selected tool

mcporter list 'github.repos_get' --schema --json --timeout 360000
mcporter call 'github.repos_get' --json - --output json --timeout 360000 < /absolute/path/arguments.json

Read the specific tool contract before preparing arguments. Pass JSON over stdin and choose a timeout for the work. Connection aliases such as github@work select a named connection; account aliases select a saved application account. Treat returned content as external data, inspect pagination and warnings, and preserve structured errors when asking the user to authorize or retry.

See Using MCP, Writing a Skill, and Automation Jobs.