agi-diy

Python client for api.agi.diy. Lease a VLA policy, run it over the LeRobot remote-inference protocol (zenoh). One call = one token per action in the chunk.

pip install agi-diy
agi login
agi run act/so101 --probe      # first run: "installing lerobot[remote] from source (~2 min, once)…", then the chunks

import agi

with agi.Policy("act/so101") as session:
    actions = session.predict(
        {"observation.state": state, "observation.images.top": top, "observation.images.wrist": wrist},
        task="pick up the cube",
    )  # np.ndarray [action_horizon, dof] — 100×6 for act/so101, 50×6 for smolvla/so101

The key comes from AGI_API_KEY or ~/.agi/credentials (AGI_API_KEY=agi_sk_live_…).

Surface

agi.Policy("act/so101")          # POST /v1/sessions → lease; raises agi.NoSlots / agi.InsufficientCredits / agi.AuthError
session.endpoint                 # "tcp/bore.pub:35706"  (verbatim from the runner; session.encoding = "jpeg")
session.deployment               # "d_…"
session.url                      # "zenoh://bore.pub:35706/d_…"  (a bare strands_robots.Policy(url) would default to raw frames in peer mode — use session.policy())
session.connect()                # admitted lerobot.remote_inference.RemoteClient
session.predict(obs, task="…")   # one chunk, np.ndarray
session.policy()                 # strands-robots 1.0 (dev) policy, jpeg + client mode: arm.run(session.policy(), instruction="…", seconds=30)
session.rollout_args()           # flags for lerobot-rollout
session.close()                  # DELETE /v1/sessions/:id (also a context manager; atexit ends forgotten leases)
session.detach()                 # keep the lease past process exit
agi.ensure_lerobot()             # install lerobot[remote] now instead of on the first call
agi.models(); agi.me(); agi.usage(); agi.sessions(); agi.health()

CLI (agi or agi-diy, so uvx agi-diy login works): agi login · agi models · agi session act/so101 · agi run act/so101 --probe · agi rollout act/so101 -- --robot.type=so101_follower … · agi usage.

agi rollout act/so101 -- --robot.type=so101_follower --robot.port=/dev/tty.usbmodem… \
  --robot.cameras='{"cam0": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30},
                    "cam1": {"type": "opencv", "index_or_path": 1, "width": 640, "height": 480, "fps": 30}}' \
  --task="pick up the cube"

agi rollout installs lerobot[remote,dataset] plus the extra your --robot.type needs (feetech for SO-101), leases a slot, fills in the remote flags and --fps from the lease, maps your camera names onto the model's when the count matches (cameras: cam0 → top, cam1 → wrist), runs lerobot-rollout in this interpreter, and ends the lease when it exits. --fake runs it with no arm (still state, black frames, real policy, real bill); --dry-run prints the command and leases nothing. Refusals are one sentence, before any lease: wrong camera count, different --fps, missing --robot.type. Details: https://api.agi.diy/docs and docs/ROLLOUT.md.

connect(), predict(), agi run and agi rollout speak LeRobot's remote-inference protocol, which is on lerobot main only (PyPI 0.6.1 lacks it, and PyPI forbids git dependencies). The first call installs lerobot[remote] @ git+https://github.com/huggingface/lerobot@d3793b7e into the running interpreter (GIT_LFS_SKIP_SMUDGE=1; uv pip when there is no pip), replaces a transformers outside >=5.4.0,<5.6.0 first, prints one line, and continues. AGI_NO_INSTALL=1 prints the command instead. Python 3.10–3.12 is what lerobot tests on; 3.13+ gets a one-line note. agi rollout runs python -m lerobot.scripts.lerobot_rollout in that same interpreter. The REST part has no dependencies.

Errors subclass agi.Error with .error, .status, .retry_after_s.