{"id":"afe952a4-000c-40da-8e2c-a66c72bb9f00","slug":"agui-framework","kind":"mcp","name":"agui-framework","summary":"A TypeScript SDK for building AI agent-powered applications. agui-framework provides a complete toolkit for creating, orchestrating, and deploying LLM-based agents with multi-provider support, real-ti","owner":"fab679","repoUrl":"https://github.com/fab679/agui-framework","license":"MIT","licenseTier":"permissive","trust":"community","stars":2,"match":null,"fewStars":true,"format":"mcp","createdAt":"2026-07-21T09:30:53.000Z","addedAt":"2026-10-21T21:33:55.815Z","logoUrl":"https://avatars.githubusercontent.com/u/23230459?u=2a56c6a626317e7114c929a518a991bc564f179f&v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["developer-tools","design-media","utilities"],"tags":[],"runtime":"node","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-25T22:42:16.616Z","sample":false,"url":"https://conduits.dev/tools/agui-framework","language":"JavaScript","transports":["stdio"],"install":{"npm":"agui-framework"},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"# agui-framework\n\n

\n \"agui-framework\"\n

\n\n

\n \n \"Documentation\"\n \n \n \"npm\"\n \n \n \"MIT\n \n

\n\nA TypeScript SDK for building AI agent-powered applications. agui-framework provides a complete toolkit for creating, orchestrating, and deploying LLM-based agents with multi-provider support, real-time streaming, state management, persistence, and the AG-UI protocol for frontend communication.\n\n## Features\n\n- **Agent class** -- Central orchestrator for LLM interactions with run/stream/resume execution modes\n- **Multi-LLM providers** -- OpenAI, Anthropic, Ollama, and Fireworks support with a common abstraction\n- **Real-time streaming** -- AsyncGenerator-based streaming with event callbacks\n- **Event system** -- Publish/subscribe EventBus with history, compaction, and piping\n- **Structured output** -- JSON Schema-enforced responses via `response_format`\n- **State management** -- Thread-isolated SharedState with versioning, diffing, merging, and conflict resolution; agents can autonomously read/write global shared state via built-in setState/getState/deleteState tools\n- **Long-term memory** -- Optional RDF-based semantic store (Oxigraph) with self-managing remember/recall/forget tools\n- **AG-UI protocol** -- Full SSE-based protocol encoding, validation, and event compaction\n- **MCP (Model Context Protocol)** -- Connect to any MCP-compatible tool server via stdio or streamable HTTP; tools are auto-discovered and registered\n- **Multi-agent patterns** -- Delegation, cyclic handoff, capability routing, and directed graph workflows\n- **Middleware pipeline** -- Composable event interception and transformation\n- **Persistence** -- Memory, Redis, and Postgres thread stores\n- **HTTP/WebSocket server** -- Express route handlers, WebSocket agent communication, model catalog API\n- **Model catalog** -- 44 models across 4 providers with pricing, context windows, and capabilities\n- **Cost & usage tracking** -- Per-run token usage, cost calculation, cumulative thread cost, budget limits\n- **WebSocket client** -- Full-duplex agent communication with run/stream/resume/capabilities\n- **React client hooks** -- useStream, useThread, useInterrupts, useCoAgent, useWebSocket, useAgentState, useRunningAgents, useLiveState, and more\n- **State REST API** -- Server endpoints for reading/writing agent shared state; AguiClient methods and React hooks included\n- **Type safety** -- Full TypeScript with strict types across all modules\n\n## Installation\n\n```bash\nnpm install agui-framework\n```\n\nRequires Node.js 18+ for native `fetch` and `ReadableStream` support.\n\nOptional persistence dependencies:\n\n```bash\nnpm install ioredis # RedisThreadStore\nnpm install pg # PostgresThreadStore\n```\n\nOptional long-term memory:\n\n```bash\nnpm install oxigraph # OxigraphSemanticStore\n```\n\n## Quick Start\n\n### 1. Basic Agent\n\n```typescript\nimport { Agent } from \"agui-framework\";\n\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"You are a helpful assistant.\",\n});\n\nconst response = await agent.run(\"What is the capital of France?\");\nconsole.log(response);\n```\n\nSet `OPENAI_API_KEY` in your environment.\n\n### 2. Streaming Response\n\n```typescript\nimport { Agent } from \"agui-framework\";\n\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"Tell a detailed story.\",\n});\n\nfor await (const chunk of agent.stream(\"Tell me about space exploration\")) {\n process.stdout.write(chunk);\n}\n```\n\n### 3. Agent with Tools\n\n```typescript\nimport { Agent } from \"agui-framework\";\nimport type { ToolConfig } from \"agui-framework\";\n\nconst weatherTool: ToolConfig = {\n name: \"get_weather\",\n description: \"Get current weather for a city\",\n parameters: {\n type: \"object\",\n properties: { city: { type: \"string\", description: \"City name\" } },\n required: [\"city\"],\n },\n handler: async ({ city }) => ({ city, temperature: 22, conditions: \"sunny\" }),\n};\n\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"Use the weather tool when asked about weather.\",\n tools: [weatherTool],\n});\n\nconst response = await agent.run(\"What is the weather in Paris?\");\nconsole.log(response);\n```\n\n### 4. Structured Output\n\n```typescript\nimport { Agent } from \"agui-framework\";\n\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"Extract structured data from text.\",\n structuredOutput: true,\n outputSchema: {\n name: \"extract_person\",\n schema: {\n type: \"object\",\n properties: {\n name: { type: \"string\" },\n age: { type: \"number\" },\n occupation: { type: \"string\" },\n },\n required: [\"name\", \"occupation\"],\n },\n },\n});\n\nconst response = await agent.run(\"Alice is a 30-year-old engineer.\");\nconsole.log(response); // JSON string: { \"name\": \"Alice\", \"age\": 30, \"occupation\": \"engineer\" }\n```\n\n### 5. Global Shared State\n\nA `SharedState` instance can be passed to one or more agents. The agent automatically registers `setState`/`getState`/`deleteState` tools that read and write the shared state — no thread scoping required.\n\n```typescript\nimport { Agent, SharedState } from \"agui-framework\";\n\nconst shared = new SharedState({ theme: \"dark\", maxRetries: 3 });\n\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"You are a helpful assistant. Use setState to remember user preferences.\",\n sharedState: shared, // enables setState/getState/deleteState tools\n});\n\nconst response = await agent.run(\"Remember my favorite color is blue.\");\n// The agent calls setState(\"favoriteColor\", \"blue\") autonomously\nconsole.log(shared.get(\"favoriteColor\")); // \"blue\"\n\n// Multiple agents or users can share the same state:\nconst agent2 = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"You know the user's preferences.\",\n sharedState: shared,\n});\n\nconst reply = await agent2.run(\"What is my favorite color?\");\n// The agent calls getState(\"favoriteColor\") and responds \"blue\"\n```\n\n### 6. MCP (Model Context Protocol)\n\nConnect any MCP-compatible tool server and use its tools automatically:\n\n```typescript\nimport { Agent } from \"agui-framework\";\n\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"You have access to MCP tools.\",\n mcpServers: [\n // Local subprocess via stdio\n { transport: \"stdio\", command: \"node\", args: [\"path/to/mcp-server.js\"] },\n // Remote server via streamable HTTP\n { transport: \"streamable-http\", url: \"https://mcp.example.com/tools\" },\n ],\n});\n\n// MCP tools are auto-discovered and merged into the agent's tool list\nconst tools = agent.getTools(); // includes MCP tools + any configured tools\n\n// Tools are available during run/stream execution\nconst response = await agent.run(\"Use the connected MCP tools.\");\n```\n\n### 7. Multi-Provider Configuration\n\n```typescript\nimport { Agent } from \"agui-framework\";\n\n// Anthropic\nconst agent = new Agent({\n model: \"claude-3-5-sonnet-20240620\",\n provider: \"anthropic\",\n instructions: \"You are helpful.\",\n});\n\n// Ollama (local)\nconst localAgent = new Agent({\n model: \"llama3\",\n provider: \"ollama\",\n instructions: \"You are helpful.\",\n baseUrl: \"http://localhost:11434/v1\",\n});\n\n// Fireworks\nconst fwAgent = new Agent({\n model: \"accounts/fireworks/models/deepseek-v4-flash\",\n provider: \"fireworks\",\n instructions: \"You are helpful.\",\n});\n```\n\n### 8. Long-Term Memory\n\n```typescript\nimport { Agent, OxigraphSemanticStore, createLTMMiddleware } from \"agui-framework\";\n\nconst store = new OxigraphSemanticStore();\nconst agent = new Agent({\n model: \"gpt-4o\",\n provider: \"openai\",\n instructions: \"You are a helpful assistant.\",\n});\n\nagent.use(createLTMMiddleware(store));\n\n// The agent autonomously calls remember/recall/forget tools\nconst reply = await agent.run(\n \"Remember I prefer short answers.\",\n { userId: \"alice\" },\n);\n```\n\n### 9. Environment-Based Configuration\n\n```env\nAGUI_PROVIDER=openai\nAGUI_MODEL=gpt-4o\nAGUI_INSTRUCTIONS=You are a helpful assistant.\nAGUI_MAX_TOKENS=1024\n```\n\n```typescript\nimport \"dotenv/config\";\nimport { Agent } from \"agui-framework\";\n\nconst agent = Agent.createFromEnv();\nconst reply = await agent.run(\"Tell me a joke\");\nconsole.log(reply);\n```\n\n## Configuration\n\n### AgentConfig\n\n| Option | Type | Default | Description |\n| ------------------------ | ------------------------- | ---------------- | --------------------------------------------------------------------------------------------- |\n| `name` | `string` | undefined | Display name for the agent |\n| `model` | `string` | required | LLM model identifier |\n| `provider` | `ProviderType` | required | `openai`, `anthropic`, `ollama`, `fireworks` |\n| `instructions` | `string` | required | System prompt |\n| `tools` | `ToolConfig[]` | `[]` | Tool definitions |\n| `maxTokens` | `number` | `1024` | Maximum output tokens |\n| `temperature` | `number` | `0.7` | Sampling temperature (0-2) |\n| `topP` | `number` | undefined | Nucleus sampling |\n| `stream` | `boolean` | `false` | Enable streaming by default |\n| `apiKey` | `string` | env var | API key for the provider |\n| `baseUrl` | `string` | provider default | Custom API endpoint |\n| `store` | `ThreadStore` | undefined | Persistence backend |\n| `autoPersist` | `boolean` | `true` | Auto-save to store after runs |\n| `capabilities` | `string[]` | `[]` | Custom capability flags |\n| `maxIterations` | `number` | `10` | Execution iteration limit |\n| `maxExecutionTime` | `number` | `30000` | Max wall-clock time in ms |\n| `structuredOutput` | `boolean` | `false` | JSON schema output support |\n| `outputSchema` | `Record` | `undefined` | JSON Schema for enforced structured output |\n| `supportedMimeTypes` | `string[]` | `['text/plain']` | Output MIME types |\n| `parallelCalls` | `boolean` | `false` | Parallel tool calls |\n| `clientProvidedTools` | `boolean` | `false` | Client-injectable tools |\n| `websocket` | `boolean` | `false` | WebSocket transport |\n| `codeExecution` | `boolean` | `false` | Declares an externally isolated code-execution capability; no in-process executor is supplied |\n| `sandboxed` | `boolean` | `false` | Metadata for an externally sandboxed tool |\n| `reasoningEncrypted` | `boolean` | `false` | Requires `reasoningEncryptionKey`; uses AES-256-GCM |\n| `reasoningEncryptionKey` | `string` | undefined | Base64-encoded 32-byte AES-256-GCM key |\n| `humanInterventions` | `boolean` | `false` | Mid-run intervention |\n| `humanFeedback` | `boolean` | `false` | Human feedback (ratings) |\n| `approveWithEdits` | `boolean` | `false` | Approve with edits |\n| `sharedState` | `SharedState` | `undefined` | Global SharedState instance; registers setState/getState/deleteState tools on the agent |\n| `mcpServers` | `MCPServerConfig[]` | `undefined` | MCP server configurations; connects on init and registers auto-discovered tools |\n\n### Environment Variables\n\n| Provider | Variable | AGUI prefix alternative |\n| --------- | ------------------- | ----------------------------------------------------------------------------------------- |\n| OpenAI | `OPENAI_API_KEY` | `AGUI_PROVIDER`, `AGUI_MODEL`, `AGUI_INSTRUCTIONS`, `AGUI_MAX_TOKENS`, `AGUI_TEMPERATURE` |\n| Anthropic | `ANTHROPIC_API_KEY` | |\n| Ollama | `OLLAMA_API_KEY` | |\n| Fireworks | `FIREWORKS_API_KEY` | |\n\n## Architecture Overview\n\n```\nagui-framework\n Agent -- Orchestrates LLM calls, tools, events, state, persistence\n EventBus -- In-process pub/sub with history and compaction\n StateManager -- Thread-isolated SharedState instances\n SharedState -- Global key-value store passed to agents via sharedState config; accessed via built-in setState/getState/deleteState tools\n MCPClientManager -- Connects to MCP servers, discovers tools, and registers them as ToolConfig entries\n ProtocolEncoder -- Event serialization and SSE encoding\n ProtocolValidator-- Input and event validation\n BaseLLMProvider -- Abstract LLM provider (OpenAI, Anthropic, Ollama, Fireworks)\n MiddlewareChain -- Composable middleware pipeline\n MultiAgentManager-- Agent delegation, cyclic handoff, capability routing, graph execution\n AgentGraph -- Directed graph of agent nodes\n ThreadStore -- Persistence interface (Memory, Redis, Postgres)\n SemanticStore -- RDF-based long-term memory interface (Oxigraph)\n LTMiddleware -- Self-managing remember/recall/forget middleware\n AguiClient -- HTTP client for remote agent execution\n React hooks -- useStream, useThread, useInterrupts, useCoAgent\n```\n\n## Documentation\n\n| Document | Description |\n| ----------------------------------------------------------------- | ------------------------------------------------- |\n| [Getting Started](docs/ag-ui-framworks-docs/getting-started.md) | Installation, prerequisites, first agent |\n| [Agents](docs/ag-ui-framworks-docs/agents.md) | Agent class deep dive: config, tools, execution |\n| [Events](docs/ag-ui-framworks-docs/events.md) | EventBus API, event types, subscriptions |\n| [State Management](docs/ag-ui-framworks-docs/state-management.md) | SharedState, StateManager, thread isolation, global agent state tools |\n| [Protocol](docs/ag-ui-framworks-docs/protocol.md) | ProtocolEncoder, SSE, validation |\n| [Providers](docs/ag-ui-framworks-docs/providers.md) | Provider architecture, custom providers |\n| [Tools](docs/ag-ui-framworks-docs/tools.md) | ToolConfig, handlers, interrupts, delegation |\n| [Architecture](docs/ag-ui-framworks-docs/architecture.md) | Module relationships, data flow, extension points |\n| [API Reference](docs/ag-ui-framworks-docs/api-reference.md) | Complete API reference organized by module |\n| [Future Integrations](docs/future-integrations.md) | Draft proposals for router, meta events, GAI |\n| [Examples](docs/ag-ui-framworks-docs/examples.md) | Multi-agent, graph, persistence, React, Express |\n| [AG-UI Agents](docs/ag-ui/Agents.md) | AG-UI protocol agent implementation |\n| [AG-UI Events](docs/ag-ui/Events.md) | AG-UI event format and types implementation |\n\n## API Overview\n\n### Agent\n\n```typescript\nclass Agent {\n constructor(config: AgentConfig);\n run(prompt: string, context?: Partial): Promise;\n stream(\n prompt: string,\n context?: Partial,\n options?: StreamingOptions,\n ): AsyncGenerator;\n resume(\n interruptId: string,\n payload?: unknown,\n status?: \"resolved\" | \"cancelled\",\n ): Promise;\n addTool(tool: ToolConfig): this;\n addCapability(capability: string): this;\n use(...middlewares: MiddlewareFunction[]): this;\n delegate(\n subAgent: Agent,\n prompt: string,\n config?: DelegationConfig,\n ): Promise;\n createDelegationTool(\n toolName: string,\n description: string,\n subAgent: Agent,\n ): ToolConfig;\n createHandoffTool(\n toolName: string,\n description: string,\n targetAgent: Agent,\n ): ToolConfig;\n clone(): Agent;\n getCapabilities(\n peerDescriptors?: Map,\n ): AgentCapabilities;\n loadThread(threadId: string): Promise;\n saveThread(threadId: string): Promise;\n toJSON(): string;\n fromJSON(json: string): this;\n static create(config: AgentConfig): Agent;\n static createFromEnv(): Agent;\n}\n\ninterface AgentConfig {\n name?: string;\n model: string;\n provider: ProviderType;\n instructions: string;\n tools?: ToolConfig[];\n maxTokens?: number;\n temperature?: number;\n topP?: number;\n stream?: boolean;\n modelSettings?: Record;\n capabilities?: string[];\n apiKey?: string;\n baseUrl?: string;\n store?: ThreadStore;\n autoPersist?: boolean;\n maxIterations?: number;\n maxExecutionTime?: number;\n structuredOutput?: boolean;\n supportedMimeTypes?: string[];\n parallelCalls?: boolean;\n clientProvidedTools?: boolean;\n websocket?: boolean;\n codeExecution?: boolean;\n sandboxed?: boolean;\n reasoningEncrypted?: boolean;\n humanInterventions?: boolean;\n humanFeedback?: boolean;\n approveWithEdits?: boolean;\n sharedState?: SharedState;\n mcpServers?: MCPServerConfig[];\n multimodalInput?: {\n image?: boolean;\n audio?: boolean;\n video?: boolean;\n pdf?: boolean;\n file?: boolean;\n };\n multimodalOutput?: { image?: boolean; audio?: boolean };\n}\n```\n\n### EventBus\n\n```typescript\nclass EventBus {\n constructor(maxHistory?: number);\n emit(event: AgentEvent): void;\n on(type: EventType | \"*\", listener: EventListener): () => void;\n once(type: EventType | \"*\", listener: EventListener): void;\n clear(): void;\n getHistory(): BaseEvent[];\n compact(): void;\n pipe(transform: (event: AgentEvent) => AgentEvent | null): EventBus;\n toJSON(): string;\n fromJSON(json: string): void;\n}\n```\n\n### StateManager / SharedState\n\n```typescript\nclass StateManager {\n constructor(defaultState?: SharedState);\n getOrCreateState(threadId: string): SharedState;\n updateState(threadId: string, updates: Partial): SharedState;\n deleteState(threadId: string): void;\n hasState(threadId: string): boolean;\n getAllThreads(): string[];\n clearAll(): void;\n exportState(threadId: string): StateSnapshot | null;\n importState(threadId: string, snapshot: StateSnapshot): SharedState;\n subscribe(subscription: StateSubscription): () => void;\n}\n\nclass SharedState {\n constructor(\n initialData?: StateData,\n options?: StateOptions,\n threadId?: string,\n );\n get(key: string, defaultValue?: T): T;\n set(key: string, value: unknown): this;\n update(updates: Partial): this;\n delete(key: string): this;\n has(key: string): boolean;\n clear(): void;\n takeSnapshot(label?: string): void;\n getSnapshot(): StateSnapshot;\n getHistory(limit?: number): StateSnapshot[];\n diff(other: SharedState): StateDiff;\n merge(\n other: SharedState,\n strategy?: StateMergeStrategy,\n conflictHandler?: Function,\n ): StateData;\n computePatch(): JsonPatchOperation[];\n getVersion(): string;\n subscribe(subscription: StateSubscription): () => void;\n toJSON(): string;\n fromJSON(json: string): this;\n}\n```\n\n### Provider Types\n\n```typescript\ntype ProviderType = \"openai\" | \"anthropic\" | \"ollama\" | \"fireworks\";\n\nabstract class BaseLLMProvider {\n readonly type: ProviderType;\n config: ProviderConfig;\n chatCompletion(\n request: ChatCompletionRequest,\n ): Promise;\n streamChatCompletion(\n request: ChatCompletionRequest,\n ): AsyncGenerator;\n prepareMessages(\n systemPrompt: string,\n messages: Array<{ role: string; content: string }>,\n ): ChatMessage[];\n}\n```\n\n### ThreadStore Interface\n\n```typescript\ninterface ThreadStore {\n connect(): Promise;\n disconnect(): Promise;\n createThread(\n threadId: string,\n metadata?: Record,\n agentId?: string,\n ownerId?: string,\n ): Promise;\n getThread(threadId: string): Promise;\n listThreads(limit?: number, offset?: number, agentId?: string, userId?: string): Promise;\n deleteThread(threadId: string): Promise;\n appendMessages(threadId: string, messages: Message[]): Promise;\n getMessages(\n threadId: string,\n limit?: number,\n offset?: number,\n ): Promise;\n saveState(threadId: string, state: Record): Promise;\n getState(threadId: string): Promise | null>;\n saveRun(runId: string, threadId: string, data: RunData): Promise;\n getRun(runId: string): Promise;\n searchMessages(\n threadId: string,\n query: string,\n limit?: number,\n ): Promise;\n}\n```\n\nImplementations: `MemoryThreadStore`, `RedisThreadStore`, `PostgresThreadStore`.\n\n### SemanticStore (Long-Term Memory)\n\n```typescript\nimport { OxigraphSemanticStore } from \"agui-framework\";\n\nconst store = new OxigraphSemanticStore();\nawait store.remember(\"alice\", [\n { subject: \"alice\", predicate: \"prefersModel\", object: \"gpt-4o\", timestamp: Date.now() },\n]);\nconst facts = await store.recall(\"alice\"); // returns all facts for alice\nawait store.forget(\"alice\", \"prefersModel\"); // delete specific fact\n```\n\nThe `createLTMMiddleware` wraps this store into the agent's execution pipeline, injecting `remember`/`recall`/`forget` tools so the agent manages its own memory. Each user's facts live in isolated named RDF graphs.\n\n### Model Catalog & Cost Tracking\n\n```typescript\nimport {\n getModel,\n calculateCost,\n formatCost,\n modelCatalog,\n} from \"agui-framework\";\n\n// Look up a model's pricing and capabilities\nconst model = getModel(\"gpt-5.6-luna\");\nconsole.log(model?.pricing); // { input: 1.00, output: 6.00, cachedInput: 0.10 }\n\n// Calculate cost from token usage\nconst cost = calculateCost(\"deepseek-v4-flash\", {\n promptTokens: 5000,\n completionTokens: 1200,\n totalTokens: 6200,\n});\nconsole.log(formatCost(cost!.totalCost)); // \"$0.0010\"\n\n// Check context window limits\nconst { exceeds, limit } = exceedsContextWindow(\"llama-3.2-3b\", 200_000);\nconsole.log(exceeds); // true, limit is 128000\n\n// Query from server\nconst client = new AguiClient(\"http://localhost:4124\");\nconst models = await client.models();\nconst gptModel = await client.model(\"gpt-5.6-terra\");\n```\n\nCosts are automatically tracked per run and accumulated per thread when a store is configured:\n\n```\nGET /api/threads → includes totalCost, runCount, lastModelId\nGET /api/threads/:id/runs → includes usage + cost per run\n```\n\n### React Hooks\n\n```typescript\n// From 'agui-framework/client/react'\nfunction useStream(): { start; stop; isLoading; error; result };\nfunction useThread(options: UseThreadOptions): {\n threads;\n messages;\n loadMessages;\n createThread;\n currentThreadId;\n};\nfunction useInterrupts(): { interrupts; handleInterrupt; resolve; clear };\nfunction useCoAgent(options: UseCoAgentOptions): UseCoAgentReturn;\nfunction useCoAction(\n toolDef: ToolConfig & { handler: Function },\n): UseCoActionReturn;\nfunction useAgentState(\n agentId: string,\n baseUrl: string,\n): { state; setState; deleteState; loading; error; refetch };\nfunction useCapabilities(\n agentId: string,\n baseUrl: string,\n): { caps; loading; error; refetch };\nfunction useAgent(\n agentId: string,\n baseUrl: string,\n): { meta; loading; error; refetch };\nfunction useWebSocket(\n baseUrl: string,\n agentId: string,\n): { connected; caps; connect; disconnect; on; off; run; resume };\n```\n\n## Event Types\n\n| Category | Events |\n| -------------- | ------------------------------------------------------------------------------------------------- |\n| Run lifecycle | `RUN_STARTED`, `RUN_FINISHED`, `RUN_ERROR` |\n| Steps | `STEP_STARTED`, `STEP_FINISHED` |\n| Text messages | `TEXT_MESSAGE_START`, `TEXT_MESSAGE_CONTENT`, `TEXT_MESSAGE_END` |\n| Tool calls | `TOOL_CALL_START`, `TOOL_CALL_ARGS`, `TOOL_CALL_END`, `TOOL_CALL_RESULT` |\n| State | `STATE_SNAPSHOT`, `STATE_DELTA`, `MESSAGES_SNAPSHOT` |\n| Activity | `ACTIVITY_SNAPSHOT`, `ACTIVITY_DELTA` |\n| Reasoning | `REASONING_START`, `REASONING_END`, `REASONING_MESSAGE_*`, `REASONING_ENCRYPTED_VALUE` |\n| Multi-agent | `AGENT_DELEGATION_START`, `AGENT_DELEGATION_END`, `AGENT_HANDOFF_REQUEST`, `AGENT_HANDOFF_RESULT` |\n| Human-in-loop | `HUMAN_INTERVENTION_REQUEST`, `HUMAN_INTERVENTION_RESULT`, `HUMAN_FEEDBACK` |\n| Code execution | `CODE_EXECUTION_START`, `CODE_EXECUTION_RESULT` |\n| Memory | `MEMORY_SUMMARY` |\n| Usage & cost | `USAGE_UPDATE` |\n| Custom | `RAW`, `CUSTOM` |\n\n## License\n\nMIT\n"}