Truth-based AI for operating systems.

Live interface

One request. A complete agent run.

Launch CommandAGI 000 · ground behind a thread with a single API call. The thread is watchable while it works and preserves the evidence behind every step.

What it returns

One endpoint — POST /v1/transduce — takes a model id and its typed inputs, and returns that model's typed outputs. GET /v1/models lists every signature.

commandagi/000-perceive: A typed scene graph from raw pixels — every window, control, and field with its box and label.
commandagi/000-ground: A phrase resolved to the one element on screen it refers to, or an honest refusal when it is genuinely ambiguous.
commandagi/000-parse-conversation: A sentence read into the typed triples of a conversation.
commandagi/000-parse-goal: A sentence read into the steps and roles of a goal.

There is no /v1/act. The model emits structure; deciding what to do with it is the policy's job — yours through the API, or ours when you run it inside a thread.

import os, requests

res = requests.post(
    "https://api.commandagi.com/threads",
    headers={"Authorization": f"Bearer {os.environ['COMMANDAGI_API_KEY']}"},
    json={
        "model": "commandagi/000-ground",
        "environment": "ubuntu",
        "task": "Open a browser and book the cheapest flight to Tokyo next month.",
    },
)
thread = res.json()
print(thread["id"], thread["watchUrl"])

CommandAGI model 000

It doesn't suggest the work. It does it.

An operating model for real computers, robots, and worlds.

Observe

Read the screen into typed, grounded objects instead of guessing from a blur of pixels.

Act

Use the real mouse, keyboard, terminal, applications, and attached machines.

Verify

Check the change every action was meant to cause, recover, and continue until done.