Letta Agent Message Streaming (HTTP + SSE)

asyncapi: '2.6.0'
id: 'urn:com:letta:api:v1:agents:messages:sse'
info:
  title: Letta Agent Message Streaming (HTTP + SSE)
  version: '1.0.0'
  description: |
    AsyncAPI 2.6 description of Letta's **agent message streaming** surface.

    Letta does not publish a WebSocket API. The only asynchronous / event-style
    transport documented in Letta's OpenAPI spec
    (https://raw.githubusercontent.com/letta-ai/letta/main/fern/openapi.json)
    is **HTTP Server-Sent Events (SSE)**, returned as `text/event-stream` from
    `POST /v1/agents/{agent_id}/messages` when the request body sets
    `stream_steps` (or the legacy, now-deprecated `POST
    /v1/agents/{agent_id}/messages/stream` endpoint). The OpenAI-compatible
    `POST /v1/chat/completions` endpoint streams the same way when
    `stream: true` is set. SSE is a one-way, server-to-client HTTP streaming
    channel; it is **not** WebSocket.

    Each SSE `data:` line carries one JSON-serialized Letta message object -
    `reasoning_message`, `assistant_message`, `tool_call_message`,
    `tool_return_message`, a periodic `ping` keep-alive, or (when
    `stream_tokens` and usage reporting are requested) a final usage
    statistics object - followed by a literal `data: [DONE]` terminator.

    This AsyncAPI document models only the streamed events. The request body
    itself (agent_id, messages, stream_steps, stream_tokens, etc.) is modeled
    in the companion OpenAPI document at `openapi/letta-ai-openapi.yml`.
  contact:
    name: API Evangelist
    email: kin@apievangelist.com
    url: https://apievangelist.com
  license:
    name: API documentation - Letta Terms of Service
    url: https://www.letta.com/terms-of-service
  x-transport-notes:
    transport: HTTP Server-Sent Events (SSE)
    protocol: https
    direction: server-to-client (one-way)
    mediaType: text/event-stream
    triggeredBy: 'POST https://api.letta.com/v1/agents/{agent_id}/messages with request body { "stream_steps": true }'
    terminator: 'data: [DONE]'
    notWebSocket: true
    source: https://raw.githubusercontent.com/letta-ai/letta/main/fern/openapi.json
defaultContentType: text/event-stream
servers:
  lettaCloud:
    url: api.letta.com/v1
    protocol: https
    description: |
      Letta Cloud's REST base. Agent message streaming is delivered as HTTP
      Server-Sent Events over this base when `stream_steps: true` (and
      optionally `stream_tokens: true`) is set on the JSON request body.
      AsyncAPI 2.6 does not define a dedicated SSE protocol identifier;
      `https` is used here and the SSE transport is documented in
      `info.x-transport-notes` and on each channel.
    security:
      - bearerAuth: []
channels:
  /agents/{agent_id}/messages:
    parameters:
      agent_id:
        description: The ID of the agent, in the format `agent-<uuid4>`.
        schema:
          type: string
    description: |
      Agent message SSE stream. The client opens this channel by issuing
      `POST /agents/{agent_id}/messages` with `Content-Type: application/json`
      and a JSON body containing `stream_steps: true` (add `stream_tokens: true`
      to additionally stream partial assistant-message tokens). The server
      responds with `Content-Type: text/event-stream` and emits an ordered
      sequence of `data:` lines, each carrying one JSON-serialized Letta
      message, followed by a final `data: [DONE]` line.
    bindings:
      http:
        type: request
        method: POST
        bindingVersion: '0.3.0'
      x-sse:
        mediaType: text/event-stream
        eventField: 'data'
        terminator: '[DONE]'
    subscribe:
      operationId: streamAgentMessages
      summary: Subscribe to streamed agent turn events (SSE).
      description: |
        After `POST /agents/{agent_id}/messages` is issued with
        `stream_steps: true`, the server emits an ordered sequence of SSE
        `data:` events representing the agent's reasoning, tool calls, tool
        results, and final assistant reply for the turn, interleaved with
        periodic `ping` keep-alives, and terminated by the literal sentinel
        `[DONE]`.
      bindings:
        http:
          type: response
          bindingVersion: '0.3.0'
      message:
        oneOf:
          - $ref: '#/components/messages/ReasoningMessage'
          - $ref: '#/components/messages/ToolCallMessage'
          - $ref: '#/components/messages/ToolReturnMessage'
          - $ref: '#/components/messages/AssistantMessage'
          - $ref: '#/components/messages/LettaPing'
          - $ref: '#/components/messages/UsageStatistics'
          - $ref: '#/components/messages/StreamDone'
components:
  securitySchemes:
    bearerAuth:
      type: http
      scheme: bearer
      bearerFormat: 'Letta API key'
      description: |
        Standard Letta bearer token, created at https://app.letta.com/api-keys.
        Set the `Authorization: Bearer <LETTA_API_KEY>` header on the
        `POST /agents/{agent_id}/messages` request that opens the SSE stream.
  messages:
    ReasoningMessage:
      name: ReasoningMessage
      title: Agent reasoning step
      summary: The agent's internal chain-of-thought for the current step, when reasoning is enabled and visible.
      contentType: application/json
      description: |
        Sent as `data: {json}\n\n` on the SSE stream. `message_type` is always
        the literal string `reasoning_message`.
      payload:
        $ref: '#/components/schemas/ReasoningMessage'
      examples:
        - name: reasoningStep
          summary: Agent reasoning before acting
          payload:
            id: message-6f7e8d9c-0000-4a1b-9c2d-3e4f5a6b7c8d
            date: '2026-07-02T18:04:11.000Z'
            message_type: reasoning_message
            reasoning: The user is asking about their prior order, I should check archival memory first.
            source: reasoner_model
    ToolCallMessage:
      name: ToolCallMessage
      title: Agent tool call
      summary: A tool invocation the agent is making as part of this step.
      contentType: application/json
      description: |
        Sent as `data: {json}\n\n` on the SSE stream. `message_type` is always
        the literal string `tool_call_message`.
      payload:
        $ref: '#/components/schemas/ToolCallMessage'
      examples:
        - name: archivalSearch
          summary: Agent calling the archival_memory_search tool
          payload:
            id: message-1a2b3c4d-0000-4a1b-9c2d-3e4f5a6b7c8d
            date: '2026-07-02T18:04:11.500Z'
            message_type: tool_call_message
            tool_call:
              name: archival_memory_search
              arguments: '{"query": "prior order"}'
              tool_call_id: call-9f8e7d6c
    ToolReturnMessage:
      name: ToolReturnMessage
      title: Agent tool result
      summary: The result returned by a tool the agent called.
      contentType: application/json
      description: |
        Sent as `data: {json}\n\n` on the SSE stream. `message_type` is always
        the literal string `tool_return_message`.
      payload:
        $ref: '#/components/schemas/ToolReturnMessage'
      examples:
        - name: archivalSearchResult
          summary: Result of the archival_memory_search call
          payload:
            id: message-2b3c4d5e-0000-4a1b-9c2d-3e4f5a6b7c8d
            date: '2026-07-02T18:04:11.900Z'
            message_type: tool_return_message
            tool_return: 'Found 1 matching passage.'
            status: success
            tool_call_id: call-9f8e7d6c
    AssistantMessage:
      name: AssistantMessage
      title: Agent reply to the user
      summary: The agent's user-facing reply content for this turn. Streamed token-by-token when stream_tokens is set.
      contentType: application/json
      description: |
        Sent as `data: {json}\n\n` on the SSE stream. `message_type` is always
        the literal string `assistant_message`.
      payload:
        $ref: '#/components/schemas/AssistantMessage'
      examples:
        - name: assistantReply
          summary: Final assistant-facing reply
          payload:
            id: message-3c4d5e6f-0000-4a1b-9c2d-3e4f5a6b7c8d
            date: '2026-07-02T18:04:12.400Z'
            message_type: assistant_message
            content: 'Your prior order shipped on June 30th.'
    LettaPing:
      name: LettaPing
      title: Keep-alive ping
      summary: A periodic keep-alive event so the SSE connection does not time out during long-running turns.
      contentType: application/json
      description: |
        Sent as `data: {json}\n\n` on the SSE stream. `message_type` is always
        the literal string `ping`. Carries no meaningful content beyond `id`
        and `date`.
      payload:
        $ref: '#/components/schemas/LettaPing'
      examples:
        - name: keepAlive
          summary: Keep-alive ping during a long-running tool call
          payload:
            id: message-ping-0001
            date: '2026-07-02T18:04:20.000Z'
            message_type: ping
    UsageStatistics:
      name: UsageStatistics
      title: End-of-turn usage statistics
      summary: Optional final chunk reporting token and step usage for the completed turn.
      contentType: application/json
      description: |
        Sent as `data: {json}\n\n` on the SSE stream, immediately before
        `data: [DONE]`, when usage reporting applies to the request.
      payload:
        $ref: '#/components/schemas/LettaUsageStatistics'
      examples:
        - name: turnUsage
          summary: Usage for the completed turn
          payload:
            message_type: usage_statistics
            completion_tokens: 128
            prompt_tokens: 512
            total_tokens: 640
            step_count: 3
    StreamDone:
      name: StreamDone
      title: Stream terminator
      summary: |
        The literal SSE event `data: [DONE]` that marks end of stream. Not
        JSON; the payload is the string `[DONE]`.
      contentType: text/plain
      description: |
        Clients must stop reading the stream when this sentinel is observed.
      payload:
        $ref: '#/components/schemas/StreamDoneSentinel'
      examples:
        - name: done
          summary: End-of-stream sentinel
          payload: '[DONE]'
  schemas:
    StreamDoneSentinel:
      type: string
      enum:
        - '[DONE]'
      description: |
        End-of-stream sentinel. The full SSE line is `data: [DONE]`. The
        payload value modeled here is the string literal `[DONE]`.
    ReasoningMessage:
      type: object
      description: The agent's internal reasoning for the current step.
      required:
        - id
        - date
        - reasoning
      properties:
        id:
          type: string
        date:
          type: string
          format: date-time
        message_type:
          type: string
          enum:
            - reasoning_message
          default: reasoning_message
        reasoning:
          type: string
        source:
          type: string
          description: Which model/pass produced this reasoning (for example, a dedicated reasoner model).
        signature:
          type: string
          nullable: true
          description: Optional provider-signed signature for the reasoning content.
    ToolCallMessage:
      type: object
      required:
        - id
        - date
        - tool_call
      properties:
        id:
          type: string
        date:
          type: string
          format: date-time
        message_type:
          type: string
          enum:
            - tool_call_message
          default: tool_call_message
        tool_call:
          type: object
          properties:
            name:
              type: string
            arguments:
              type: string
              description: JSON-encoded arguments the model generated for the call.
            tool_call_id:
              type: string
    ToolReturnMessage:
      type: object
      required:
        - id
        - date
        - tool_return
        - status
        - tool_call_id
      properties:
        id:
          type: string
        date:
          type: string
          format: date-time
        message_type:
          type: string
          enum:
            - tool_return_message
          default: tool_return_message
        tool_return:
          type: string
        status:
          type: string
          enum:
            - success
            - error
        tool_call_id:
          type: string
        stdout:
          type: array
          items:
            type: string
        stderr:
          type: array
          items:
            type: string
    AssistantMessage:
      type: object
      required:
        - id
        - date
        - content
      properties:
        id:
          type: string
        date:
          type: string
          format: date-time
        message_type:
          type: string
          enum:
            - assistant_message
          default: assistant_message
        content:
          type: string
          description: The agent's user-facing reply text (or a token fragment, when stream_tokens is set).
    LettaPing:
      type: object
      required:
        - id
        - date
      properties:
        id:
          type: string
        date:
          type: string
          format: date-time
        message_type:
          type: string
          enum:
            - ping
          default: ping
          description: Keep-alive event to prevent SSE streams from timing out during long-running requests.
    LettaUsageStatistics:
      type: object
      description: Token and step usage for a completed agent turn.
      properties:
        message_type:
          type: string
          enum:
            - usage_statistics
        completion_tokens:
          type: integer
        prompt_tokens:
          type: integer
        total_tokens:
          type: integer
        step_count:
          type: integer
          description: Number of model-call steps the agent took to complete the turn.
        cached_input_tokens:
          type: integer
        reasoning_tokens:
          type: integer