Debug any agent in real-time.

A workbench, not a log viewer

Everything you'd otherwise wire up by hand (request builder, live stream, replay, diffing) in one place that speaks agent, not HTTP.

When your agent delegates, you can still see everything

Most tools flatten a spawned subagent into a single opaque tool call. Langport reconstructs the tree from the event stream (parent run, child runs, and every hop below), so a nested delegation reads like the conversation it actually was.

If it streams events, it works

Langport speaks plain SSE and NDJSON. Emit whatever shape you already emit: it maps tool calls, deltas, and parent ids on the way in. Point it at your agent and hit send.

# Nothing to install. Just stream events.

async def run(prompt: str):
    yield sse("token", {"text": "Splitting into 3 readers…"})

    for batch in batches:
        yield sse("tool_call", {
            "id": call_id,
            "name": "spawn_agent",
            "args": {"agent": "reader", "task": batch.prompt},
        })

    # child events carry the parent's call id
    yield sse("token", {
        "text": chunk,
        "parentToolCallId": call_id,
    })

docker run -p 3000:80 ghcr.io/langport-dev/langport:latest