You focus on building robots.
We give you the brain over cloud.
Open-weight VLA policies, served over Zenoh with LeRobot's remote-inference protocol.
Your robot, your loop, our GPU.
1 action = 1 token = $0.0001. $20 on signup.
pip install agi-diy
agi login
agi run act/so101 --probe● act/so101 · smolvla/so101 · SO-101 arm
how it works
- login — GitHub, one click, $20 of credit on the account.
- key —
agi_sk_live_…, shown once, used asAuthorization: Beareroragi login. - lease —
POST /v1/sessions {"model":"act/so101"}returns a Zenoh endpoint and a deployment name. - zenoh — your LeRobot or strands-robots client connects unchanged; observations go up, action chunks come back.
measured
2026-10-24, from a Mac on another network, through the public relay:
act/so101 · 100 actions per call · 62–316 ms
smolvla/so101 · 50 actions per call · 414–835 ms
what this is
- A proof of concept. One Jetson Thor, two models, five slots.
- Your robot stays yours. Observations go up, action chunks come back. Nothing else.
- Keys are shown once. We keep a hash.
- LeRobot and strands-robots clients work unchanged. You get an endpoint and a deployment name.
what this isn't
- Fine-tuning. Not yet.
- An SLA. One runner, one uplink, no redundancy.
- Bimanual or mobile. SO-101 only today.
- More than 5 sessions at once.