The Confab API

Your first request

1. Make a key in the Confab Console. You see it once, so copy it then.
2. Put it in an environment variable called CONFAB_API_KEY, not in your code.
3. Send a message:

curl https://api.confab.im/v1/messages \
  -H "Authorization: Bearer $CONFAB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "confab-core",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Hello, Confab" }]
  }'

import os
import requests

response = requests.post(
    "https://api.confab.im/v1/messages",
    headers={"Authorization": f"Bearer {os.environ['CONFAB_API_KEY']}"},
    json={
        "model": "confab-core",
        "max_tokens": 1024,
        "messages": [{"role": "user", "content": "Hello, Confab"}],
    },
)
print(response.json()["content"][-1]["text"])

Requests and answers use the widely used messages format: a system prompt, the conversation, max_tokens and content blocks. A client library made for it works once its base URL is https://api.confab.im and it has your Confab key.

Two models

How they perform

Measured on the machines that serve the API while they weren’t busy: chat on 31 October 2026, pictures on 24 October. Your own network adds to the times.

Chat: Qwen3.8 27B answering one request at a time, about 20 tokens a second for prose and 34 for lists and code, and reading an 11,000-token conversation at about 850 a second. Confab Image: real requests for one picture, safety check included; the first after a quiet spell can take longer while the model loads.

The Confab look, tested

We made 100 new pictures both ways, in the Confab look and as described, with the same prompt and seed each time, and scored them with OpenAI’s CLIP and LAION’s aesthetic predictor. In the Confab look they came out:

• Closer to the house style in 76 of 100 pictures (CLIP similarity 0.778, against 0.744).
• A slightly better match for the prompt (CLIP score 26.9, against 26.2).
• Just as bright, and a little less saturated (22.9% colour saturation, against 25.0%).
• Scored a little lower by LAION's aesthetic predictor (5.41, against 5.62), which prefers vivid, polished pictures to calm, muted ones.

Uptime for every part of Confab, checked every minute, is on the status page.

Everything in the Confab Console

API keys

Make keys for each app, choose what each can do, set an expiry, and rotate one without downtime.

Playground

Try the chat model and Confab Image with your settings, then copy the request as curl, Python or TypeScript.

Logs

Every request for 30 days, with the full request and response for 7, and what to do about an error.

Usage

Requests, tokens and pictures by day, model and key, your busiest minute against each limit, and a CSV export.

The full reference (authentication, streaming events, errors and the usage and logs endpoints) is in the Console docs.

Limits in the beta

Per account, shared by all its keys. A request over a limit gets a 429 with a retry-after header.

Free in beta

Requests cost nothing while the API is in beta. When pricing starts, they are paid from your Confab balance, and a monthly spend limit can stop them before they go past it. See pricing