deepagents-serve 0.1.0

Universal agent server. Wraps any LangGraph graph and exposes it over standard agent protocols — Vercel AI SDK, Claude Code, TUIs, and more.

deepagents-serve

Universal agent server. Wraps any LangGraph graph and exposes it over standard agent protocols — Vercel AI SDK, Claude Code, TUIs, and more.

Quick start

from deepagents_serve import DeepAgentsServeApp
from deepagents import create_deep_agent
import uvicorn

agent = create_deep_agent(
    model="anthropic:claude-opus-4-8",
    skills=["myskillsregistry/skills"],
    system_prompt="You are a helpful assistant."
)

app = DeepAgentsServeApp(agent)

uvicorn.run(app.build(), host="0.0.0.0", port=8000)

---

Endpoint design

POST /chat/stream

Stateless. One request, one SSE stream, done. No session state is kept between calls.

Request

{
  "messages": [
    {"role": "user", "content": "List the files in the working directory."}
  ]
}

Response — Content-Type: text/event-stream

Uses the Vercel AI Data Protocol. Each line is data: <json>:

data: {"type": "text-start", "id": "msg_01"}
data: {"type": "text-delta", "id": "msg_01", "delta": "Here are the files:\n"}
data: {"type": "text-end", "id": "msg_01"}

data: {"type": "tool-input-start", "toolCallId": "tc_01", "toolName": "bash"}
data: {"type": "tool-input-delta", "toolCallId": "tc_01", "inputTextDelta": "{\"command\":\"ls\"}"}
data: {"type": "tool-output-available", "toolCallId": "tc_01", "output": "README.md\nsrc/"}

data: {"type": "reasoning-start", "id": "r_01"}
data: {"type": "reasoning-delta", "id": "r_01", "delta": "I should list the files..."}
data: {"type": "reasoning-end", "id": "r_01"}

data: {"type": "error", "errorText": "something went wrong"}
data: {"type": "abort", "reason": "user cancelled"}
data: {"type": "finish"}

---

Signal diagram

Client (Vercel SDK / Claude Code / TUI)
  │
  │  POST /chat/stream
  │  {"messages": [{"role": "user", "content": "..."}]}
  ▼
┌──────────────────┐
│  Input Parser    │  extract last user message from messages array
└────────┬─────────┘
         │ {"role": "user", "content": "..."}
         ▼
┌──────────────────┐
│  LangGraph Graph │  graph.astream({"messages": [...]}, stream_mode=["messages"])
└────────┬─────────┘
         │ (stream_part, metadata) chunks
         ▼
┌──────────────────┐
│  Normaliser      │  LangGraph messages → canonical AgentEvent
│                  │
│  AIMessage       │→  TextDelta | ReasoningDelta
│  ToolCall        │→  ToolInputDelta
│  ToolMessage     │→  ToolOutputAvailable
│  Error           │→  Error | Abort
└────────┬─────────┘
         │ AgentEvent
         ▼
┌──────────────────┐
│  Vercel Adapter  │  AgentEvent → Vercel SSE stream parts
└────────┬─────────┘
         │ text/event-stream
         ▼
Client

---

Roadmap

Near-term

• Input parsing — read request body, extract messages
• Complete normaliser — tool calls, tool results, reasoning deltas from LangGraph
• Request validation — reject malformed payloads at the boundary

Protocol compatibility

• Claude Managed Agents endpoints — /v1/sessions, /v1/sessions/{id}/events, /v1/sessions/{id}/events/stream
• Claude Managed Agents SSE encoding — agent.message, agent.tool_use, session.status_idle event types
• Agent metadata endpoint — GET /v1/agents/default

Statefulness

• Session layer — LangGraph thread_id checkpointing, session_id → thread_id mapping
• Multi-turn conversations — persist conversation history across requests
• Interrupt support — user.interrupt event stops graph mid-execution

Deployment

• kagent packaging — Kubernetes-compatible worker pattern

Eval

• Event history capture — persist full turn events with timestamps and token counts
• Session replay — re-run a captured session for regression testing

Skills

• Skill registry compatibility — discover and invoke skills from a registry
• Finetune on skills — export session history as structured training data