Framework integrations

MCP tools

@openagentid/mcp exports withOpenAgent, withOpenAgentClient, buildIdentityMeta, skills-policy helpers, and typed authentication/authorization errors. Configure an identity verifier, expected tool scopes, and the application's policy hooks. Identity metadata from a client is untrusted until verified.

The wrapper operates on the package's McpServerLike/McpClientLike contract. Match the actual MCP SDK version and handler API you deploy. Test anonymous calls, missing identity, failed verification, denied scope, and denied skills policy. The documentation MCP endpoint is a separate read-only service and is not this runtime authentication integration.

Claude Agent SDK hooks

@openagentid/claude-agent exports openAgentPlugin, lifecycle/tool/skill hook functions, capability and skills-policy implementations, and audit-chain utilities. Connect preToolUse and skill decisions before execution and postToolUse after the result. A recorded audit event does not retroactively prevent an unauthorized tool call.

Use the package's plugin/hook types to bridge the exact host SDK version. The package's source interface is the reference; do not assume that any arbitrary object in a host's plugins array has equivalent behavior. Configure real identity/capability providers and a durable audit sink where required.

Scaffold and test tooling

packages/create-openagent produces project scaffolding. Inspect its templates and generated configuration before treating a project as runnable. The separate testing SDK and conformance runners supply explicit fixtures and tests; they are not production identity providers.

All integration package manifests, exported entry points, errors and supported source paths are listed in the catalog. Pair them with skills policy, ACT verification, and authority verification.