CopilotKit Intelligence skills for Mastra

Keep published skills from one or more Learning containers available to Mastra agents. The adapter adds a catalog and two native read tools. Each invocation uses one verified snapshot, including tool-first resumes.

Requires Node.js 22.13+ and Mastra core >=1.0.0 <2. The Intelligence server and canonical Runtime client must support the learned-snapshot API.

Setup

import { Agent } from "@mastra/core/agent";
import {
  SkillRegistry,
  createSkillRegistryProcessor,
} from "@copilotkit/intelligence-mastra";

// Set CPK_INTELLIGENCE_API_KEY and CPK_INTELLIGENCE_LEARNING_CONTAINER_ID.
// Set INTELLIGENCE_API_URL for a self-hosted Intelligence server.
const registry = new SkillRegistry();
await registry.initialize();
const skills = createSkillRegistryProcessor({ registry });
const agent = skills.wrapAgent(
  new Agent({
    id: "support",
    name: "Support",
    model: "openai/gpt-4.1",
    instructions: "Follow our support policy.",
    inputProcessors: [skills],
    tools: { ...skills.tools },
  }),
);

const result = await agent.generate("Help with a refund.");

Register other processors and tools in the same native arrays and maps. Reserve copilotkit_load_skill and copilotkit_read_skill_file for this adapter. Both tools remain available for an empty registry. Host instructions retain precedence over learned content.

Execution and resume

Call the wrapped Agent for generate, stream, resumeGenerate, and resumeStream. The wrapper also covers the native approval helpers: approveToolCall, declineToolCall, approveToolCallGenerate, and declineToolCallGenerate where the installed Mastra version provides them.

Each call acquires its snapshot before native execution. Streaming results retain Mastra's native shape. A resume starts a new invocation and captures the current snapshot. Private async context carries that snapshot, so restored tools cannot read an older snapshot from a suspended closure. Snapshots are not stored in request context or processor state.

Pass abortSignal in the invocation options to cancel the delivery wait and native execution. A signal supplied only through Agent defaultOptions takes effect during native execution and cannot cancel the preceding delivery wait. Cancelling one caller does not cancel a registry refresh shared with other callers. Register and wrap each selected agent explicitly; subagent propagation follows Mastra. Agent networks, legacy methods, background workers, and separate durable-worker dispatch are outside this adapter's supported entry points.

Configuration

Pass an application-owned CopilotKitIntelligence client from @copilotkit/runtime/v2 to new SkillRegistry({ client, containerId }) to reuse its credentials and HTTP transport. Explicit configuration overrides environment values. An injected client is authoritative for connection configuration.

freshnessWindowMs and requestTimeoutMs default to 5000. debug defaults to false. Set revision or CPK_INTELLIGENCE_SKILLS_REVISION for an exact whole-container pin. Explicit initialization is optional; the first wrapped invocation can initialize the registry.

Catch SkillDeliveryError during initialization or invocation. Its stable code describes the failure. Warm transient errors retain the previous verified snapshot; confirmed denial blocks new invocations. registry.status exposes initialization, revision, mode, freshness, and safe error status.

The adapter reads skills in memory and does not write files or execute scripts. The model chooses whether to load or follow skills. The package uses the canonical Runtime client and retains that client's dependency footprint; it does not depend on the LangGraph adapter.

See the learned skill delivery guide for server requirements and migration from CLI downloads.

Multiple containers

Use the same registry with an explicit list:

const registry = new SkillRegistry({
  containers: [
    { id: "support", revision: "revision-123" },
    { id: "company-wide" },
  ],
});

Each entry follows latest unless it has a revision pin. IDs must be unique and nonempty. The SDK rejects an empty list or a list combined with containerId or top-level revision. TypeScript rejects mixed forms at compile time too. The old interface remains supported. An explicit list ignores legacy container and revision environment variables. Credentials, freshness, and timeouts remain shared.

The new interface uses names such as support/refund-policy in the catalog and tool calls. The prefix URI-encodes the container ID, so names cannot collide between containers. The old interface keeps unprefixed Skill names.

Each container keeps its own cache and revision. Every invocation captures the full combined catalog. A cold failure or confirmed denial from any container blocks the invocation. Existing transient-error fallback applies separately to each warm container. registry.status.containers lists per-container status and revision. Aggregate status has no revision pin.

Explicit containers accepts 1–50 unique container IDs and sends one batch request for all sources that need a refresh. This also applies to a list with one entry. The server must support POST /api/v1/learning/skills/batch before you use this configuration. The SDK does not fall back to separate requests. Legacy containerId configuration keeps its existing single-container request. Both interfaces use the same authentication configuration.