smooai-smooth-operator-server 1.27.0

Native async WebSocket server for the smooth-operator protocol — parity with the Rust and C# reference servers, consuming the in-process smooai-smooth-operator-core engine.

smooai-smooth-operator-server

Wiring a chat loop is a weekend project. A production agent server is not.

Sessions that survive a reconnect. A wire protocol your clients can actually speak. Streaming turns you can watch token by token. Tools the model can call — and hard limits on the ones it must never call. Human-in-the-loop when a tool wants to write.

smooai-smooth-operator-server is that server, async and native to Python. It speaks the smooth-operator wire protocol (spec/) and consumes the in-process smooai-smooth-operator-core engine — each turn runs a SmoothAgent and maps its stream onto stream_token / stream_chunk / eventual_response. It's the Python sibling of the Rust, Go, TypeScript, and C# servers, all speaking the one protocol.

The client lives in python/src (smooai-smooth-operator). This is the server half.

---

Spin up a real agent server

python -m smooth_operator_server
# → smooth-operator-server (local flavor, python) listening on ws://127.0.0.1:8787/ws

That's a full agent backend — sessions, streaming turns, tool-calling, citations — on one WebSocket, in-memory, auth off, zero config. Env knobs: SMOOTH_OPERATOR_BIND (default 127.0.0.1:8787), SMOOTH_OPERATOR_SEED_KB=1 for the demo knowledge docs. The gateway is read from SMOOAI_GATEWAY_URL / SMOOAI_GATEWAY_KEY — with no key, send_message returns a clean LLM_UNAVAILABLE error and the rest of the protocol still works.

Or embed it in your own async app:

import asyncio
from smooth_operator_server import ServerState, serve
from smooth_operator_server.session_store import InMemorySessionStore

async def main():
    state = ServerState(store=InMemorySessionStore(), chat_client=my_openai_client)
    server = await serve(state, "127.0.0.1", 0)  # port 0 → ephemeral
    print(server.ws_url())
    # ... drive real streaming turns ...
    await server.shutdown()  # graceful drain + clean exit

asyncio.run(main())

---

Extensible — and safe by construction

An agent is only useful when it can do things, and only trustworthy when you can say what it may never do. This server gives you both seams.

Give it your tools. Hand engine Tools to ServerState and they merge with the built-ins for every turn:

from smooth_operator_core import Tool  # the engine's tool base

class OpenTicket(Tool):
    name = "open_ticket"
    description = "Open a support ticket for the current customer."
    parameters = {"type": "object", "properties": {"subject": {"type": "string"}}}

    async def execute(self, arguments):
        return f"ticket opened: {arguments}"

state = ServerState(
    store=InMemorySessionStore(),
    chat_client=my_client,
    tools=[OpenTicket()],
)

Or let it gain tools with no redeploy. The server hosts SEP extensions — out-of-process tool providers discovered at runtime, their ui/confirm prompts bridged into the protocol's confirmation frames for HITL. Gated: an extension contributes tools only if you name it in SMOOTH_EXTENSIONS_ALLOW. Nothing loads by default.

Now declare the lines it can't cross. Install an AgentConfigResolver, and every tool — built-in, yours, or from an extension — flows through the same gates:

• Per-agent allow-list — an agent's tool_config.enabledTools restricts its turn to exactly those tools. Off the list, off the table.
• The authLevel gate — a tool declaring supports_auth_requirement = True is blocked at call time on a public agent when tagged admin, or when tagged end_user and the session isn't identity-verified (its OTP bit, or your SessionAuthenticator seam). Fail-closed — an absent authenticator means "not authenticated".
• End-user OTP flow — a refused end_user tool can offer a one-time-code identity flow via the OtpService seam; the server never generates, delivers, or validates a code (the host owns generation, delivery, expiry, attempt counting).

You decide what the agent can touch; the runner enforces it.

---

What it does

Piece	Module	Mirrors
WS transport + per-connection loop + single writer	server.py	Rust server.rs, C# SmoothOperatorWebSocketExtensions
Frame dispatch (ping / create / get / send_message)	dispatcher.py	C# FrameDispatcher, Rust handler.rs
Session + message store	session_store.py	C# SessionStore, Rust storage adapter
Streaming turn (engine → protocol events)	turn_runner.py	C# TurnRunner, Rust runner.rs
Per-agent config (instructions / workflow / persona / tools)	agent_config.py	monorepo agents schema
Conversation-workflow steps + post-turn judge	workflow.py	monorepo general-agent/workflow.ts
SEP extension hosting	extensions.py	Rust extensions.rs
Auth verifier seam (permissive + local HS256 JWT)	auth.py	C# Auth.cs, Rust verifier seam

Graceful SIGTERM drain. A shared asyncio.Event cancel switch is the single source of truth for "stop". Each connection loop races "cancel set" vs "next inbound frame" — with the turn dispatch awaited inside the frame branch, so an in-flight turn finishes before the loop exits, then a backplane detach always runs.

Per-agent config + conversation workflows. create_conversation_session carries only an agent UUID, so config is resolved server-side per turn: an agent's instructions become its system prompt; personality / greeting are appended (greeting only on the first turn); a conversation_workflow (goal + ordered steps) renders the current step into the prompt and a cheap post-turn judge advances the pointer when the criteria are met. Parsing is tolerant (malformed → server default, never crashes a session) and the judge is failure-tolerant (any error → stay on the current step). With no resolver installed, behavior is unchanged.

---

Five languages, one protocol

The same server — same wire protocol, same conformance corpus — exists in five languages. Run it where your stack already lives.

Language	Server package	Registry
Python	smooai-smooth-operator-server	in-repo (this package)
Rust	smooai-smooth-operator-server	crates.io
C# / .NET	SmooAI.SmoothOperator.Server	in-repo
TypeScript	@smooai/smooth-operator-server	in-repo
Go	github.com/SmooAI/smooth-operator/go/server	in-repo

Every native client — TypeScript, Go, .NET, Python, Rust — connects to any of them unmodified.

---

Develop

cd python/server
uv sync
uv run --quiet ruff format .
uv run --quiet ruff check .
uv run --quiet pytest -q

---

Part of the smooth-operator service — Smoo AI's polyglot AI agent service. Don't want to run it yourself? lom.smoo.ai hosts it for you.

License

MIT © 2026 Smoo AI. See LICENSE.