Switchboard Open workbench ↗
A PERSON ON THE OTHER END

Connect your app.
Let a human take the next turn.

Switchboard speaks familiar LLM APIs. Instead of generating an answer, it holds your request until an invited person claims it in the browser.

Two ways to join

Bring a request

Get a client API key from the administrator. Point your OpenAI or Anthropic client at this instance and choose the human model.

Your key belongs to a tenant. Only clients in that tenant can read its requests, files, conversations and usage.

Take a seat

Get a workbench invitation from the administrator, then sign in with your name and workbench key.

Claim a request, read the context, publish useful progress and leave a clear handoff. Workbench access is for trusted operators who may read all tenants’ requests.

Connection settings

Hosted origin: https://switchboard.freifick.net

ClientBase URLEndpoints
OpenAIhttps://switchboard.freifick.net/v1Chat, Completions, Responses, Conversations, Files, Batches
Anthropichttps://switchboard.freifick.netMessages, token counting, Message Batches

Use Authorization: Bearer YOUR_CLIENT_KEY or x-api-key. Keep the key in your client’s environment or credential store. Workbench keys and client keys have different roles; keep both private.

import os
import httpx
from openai import OpenAI

client = OpenAI(
    base_url="https://switchboard.freifick.net/v1",
    api_key=os.environ["SWITCHBOARD_API_KEY"],
    timeout=httpx.Timeout(None, connect=10),
    max_retries=0,
)

with client.responses.stream(
    model="human", input="Review my plan."
) as stream:
    for event in stream:
        if event.type == "response.output_text.delta":
            print(event.delta, end="", flush=True)

Waiting is part of the workflow

A person may arrive minutes or days later. Disable the client’s read timeout. Streaming sends immediate headers and periodic heartbeats while the request waits. Published answer chunks arrive immediately; a private draft stays in the workbench.

Prefer not to hold a connection? Use Responses with background: true, or send Prefer: respond-async to receive a request receipt. Poll it later, or reconnect to its events URL with Last-Event-ID. A disconnected caller does not cancel the request.

Use a stable Idempotency-Key when retrying one request. The default deadline is seven days; human.request_ttl_seconds: 0 permits an indefinite wait.

Tools, context and accounting

Describe your tools using the usual API schema. The worker gets a form for the arguments. Your client receives the tool call, executes it and sends back the result as a new turn. A shared summary, tasks and notes help the next worker understand the session.

Structured outputs, reasoning summaries, files, native items and measured usage can be authored in the workbench. Token usage is explicitly labelled when estimated. Configured token rates and claimed human time produce an accounting ledger; the default rates are zero and this instance collects no payments.

Before sending private material

Invited workers can read your request and attachments. Send only material they are permitted to see. Terminal requests and events are normally retained for 30 days; conversation context, uploaded files and accounting remain until removed by policy. Backups have a separate 30-day retention period.

Full API contract · Operations & access · Source & MIT license · Readiness