Open Think Start Free →

This document defines the capability contract used by AgentPlugin.capabilities.

§Allowed capabilities

  • tools — performs operational tool-style actions.

  • mcp — integrates with MCP server workflows (inbound or outbound).

  • cloudflare-api — performs Cloudflare API operations.

  • connectors — talks to external providers/connectors.

  • models — model/provider-specific model operations.

  • artifacts — Cloudflare Artifacts repository operations.

  • admin — runtime introspection and setup control (reserved for the admin plugin).

Runtime validation enforces:

  1. Plugin IDs must be unique in the registry.

  2. Each plugin must declare at least one capability.

  3. Every declared capability must be in the allowed set above.

§First-party plugin mapping

Plugin IDCapabilitiesBinding / secret required
adminadmin, toolsruntime introspection handle (provided automatically)
cloudflare-api-mcpcloudflare-api, mcp, toolsCLOUDFLARE_API_TOKEN or CLOUDFLARE_AGENT_TOKEN
workers-aimodels, toolsenv.AI binding
anthropicmodels, connectorsANTHROPIC_API_KEY
openai-compatiblemodels, connectorsOPENAI_COMPATIBLE_URL (+ optional key)
mcp-clientmcp, tools, connectorsMCP_DEFAULT_URL (or per-call serverUrl)
browsertools, connectorsenv.BROWSER binding
sandboxtools, connectorsenv.SANDBOX service binding
artifactsartifacts, toolsenv.ARTIFACTS binding
mppmodels, connectorsMPP_API_KEY

§Guidance for contributors

  • Choose the smallest capability set that matches behavior.

  • Do not mark capabilities "just in case".

  • If a new capability is truly needed, update:

    • src/core/plugin.ts union type
    • src/core/runtime.ts capability validator set
    • this document