Deploying a Flow as a Service

Where we left off

Every REST call so far (Lessons 7, 8, 11, 16) went to the full langflow run server, the same one hosting the canvas, the Playground, account management, everything. lfx serve is smaller: point it at a flow.json, get back a standalone API server for exactly that flow and nothing else, no UI, no canvas, no login.

Why this matters

The full Langflow server is what you build and iterate on a flow with. It's not what you'd actually want running in production for a flow that's done, a UI server is a lot of surface area (and dependencies) to carry just to answer POST /run. lfx serve is the deployment-shaped version: one flow (or a folder of them), one lightweight process, the same shape a small internal service or a Docker container running a single flow would actually look like.

Do this yourself

export LANGFLOW_API_KEY=<the same key from Lesson 7>
uv run lfx serve lessons/langflow/01_beginner/02_first_flow/flow.json --port 8123

Run this in its own terminal, leave it running. In a second terminal:

curl http://127.0.0.1:8123/flows -H "x-api-key: $LANGFLOW_API_KEY"
curl -X POST http://127.0.0.1:8123/flows/<flow_id>/run \
  -H "x-api-key: $LANGFLOW_API_KEY" -H "Content-Type: application/json" \
  -d '{"input_value": "Say hello in exactly three words."}'

Notice the response shape: {"result": "...", "success": true, ...}, different from the full server's outputs[].outputs[].messages[] nesting from Lesson 11. lfx serve is a separate, smaller API surface, not a mirror of the full one.

The code, piece by piece

subprocess.Popen(
    [sys.executable, "-m", "lfx", "serve", str(FLOW_PATH), "--port", str(PORT)],
    ...
)

The same command as "Do this yourself," just started from Python so lesson.py stays a single self-contained script, in a real deployment you'd run the lfx serve ... command directly instead.

for _ in range(30):
    try:
        httpx.get(f"{SERVE_URL}/flows", headers=HEADERS, timeout=1)
        return process
    except httpx.TransportError:
        time.sleep(1)

The server takes a moment to start up, this polls until it actually answers instead of guessing a fixed sleep duration.

response = httpx.post(f"{SERVE_URL}/flows/{flow_id}/run", ...)
print(f"Result: {response.json()['result']}")

Same idea as run_flow() from Lesson 11, flows/{flow_id}/run instead of api/v1/run/{flow_id}, and the response's answer lives under result, not nested several levels deep.