strands-osmo

NVIDIA OSMO workflow orchestration for Strands Agents - submit, monitor, and debug Physical AI pipelines from natural language.

OSMO is NVIDIA's open-source Kubernetes-native control plane for Physical AI. It runs heterogeneous compute (training GPUs, simulation GPUs, edge devices) from a single YAML spec - used in production for GR00T, Isaac Lab, Isaac Dexterity, Isaac Sim, Isaac ROS.

This package wraps the production osmo CLI as 22 Strands tools so an agent can drive the entire workflow lifecycle in plain English.

Sister project: strands-cosmos (Cosmos VLM provider + 21 generation/edge tools).

---

🔌 Use as an MCP server

Use strands-osmo from Claude Code, Claude Desktop, Cursor, Kiro, or any MCP client — all 25 OSMO tools (workflows, tasks, data, datasets, apps, pools) become MCP tools.

claude mcp add osmo -- uvx strands-osmo

Claude Desktop config:

{
  "mcpServers": {
    "osmo": {
      "command": "uvx",
      "args": ["strands-osmo"]
    }
  }
}

Options:

strands-osmo --tools osmo_workflow_submit,osmo_workflow_status   # subset
strands-osmo --http --port 8000                                  # HTTP mode

---

Install

pip install strands-osmo

You also need the OSMO CLI on PATH:

curl -fsSL https://raw.githubusercontent.com/NVIDIA/OSMO/main/install.sh | bash
osmo login                  # one-time OAuth
osmo version                # confirm

Verify everything from your agent:

from strands_osmo import osmo_doctor
osmo_doctor()
# → {"osmo_present": True, "version_ok": True, "logged_in": True, ...}

---

Quick Start

Discover resources, submit a workflow

from strands import Agent
from strands_osmo import (
    osmo_doctor, osmo_pool_list, osmo_resources_list,
    osmo_workflow_validate, osmo_workflow_submit, osmo_workflow_status,
    osmo_workflow_logs, osmo_workflow_cancel,
    osmo_cookbook_fetch,
)

agent = Agent(tools=[
    osmo_doctor, osmo_pool_list, osmo_resources_list,
    osmo_workflow_validate, osmo_workflow_submit, osmo_workflow_status,
    osmo_workflow_logs, osmo_workflow_cancel,
    osmo_cookbook_fetch,
])

agent("""
Find me an idle H100 pool with at least 4 GPUs free,
fetch the GR00T fine-tune recipe from the cookbook,
validate it, then submit it. Report the workflow ID.
""")

One-shot pipeline

from strands_osmo import osmo_workflow_submit

osmo_workflow_submit(
    workflow_yaml="train.yaml",
    pool="h100-prod",
    set_vars={"num_gpu": 4, "epochs": 10},
    priority="LOW",   # bypass quota, run on idle capacity
)

---

Tools

Category	Tools	Wraps
Auth	osmo_doctor, osmo_login, osmo_version	env / osmo login / osmo version
Resources	osmo_pool_list, osmo_resources_list, osmo_profile_list, osmo_bucket_list	osmo pool/resources/profile/bucket list
Workflow	osmo_workflow_submit, osmo_workflow_list, osmo_workflow_status, osmo_workflow_cancel, osmo_workflow_logs, osmo_workflow_exec	osmo workflow ...
Task	osmo_task_list, osmo_task_describe	osmo task ...
Data	osmo_data_upload, osmo_data_download, osmo_data_list	osmo data ...
Dataset	osmo_dataset_list, osmo_dataset_describe	osmo dataset ...
App	osmo_app_list, osmo_app_create	osmo app ...
Cookbook	osmo_cookbook_fetch	GitHub raw / local clone
Spec	osmo_workflow_validate, osmo_workflow_render	local YAML + Jinja

Total: 25 tools (22 OSMO-CLI-backed + 3 local helpers).

from strands_osmo import osmo_pool_list, osmo_workflow_submit

osmo_pool_list(mode="free")
# → JSON: pools, free GPUs, quota state

osmo_workflow_submit("train.yaml", pool="h100-prod", set_vars={"epochs": 10})
# → workflow ID(s)

---

Why use it inside an agent?

OSMO already has a great CLI. The point of strands-osmo is letting an agent chain OSMO ops with reasoning:

Without an agent	With an agent
osmo pool list → eyeball	"Find me an idle H100"
Edit YAML by hand to fit quota	"Cap memory to fit the smallest A100 node, then submit"
osmo workflow logs <id> → grep for stack trace	"Why did workflow X fail and how do I fix it?"
osmo workflow submit && watch status	"Submit, wait until done, summarize results, retrain on failure"

---

Architecture

strands_osmo/
├── __init__.py             # exports 25 tools
└── tools/
    ├── _common.py          # osmo CLI runner + ToolResult helpers
    ├── doctor.py           # env probe (osmo binary, login, profile)
    ├── login.py            # osmo login
    ├── version.py          # osmo version
    ├── pool_list.py        # osmo pool list [--mode free]
    ├── resources_list.py   # osmo resources list
    ├── profile_list.py     # osmo profile list
    ├── bucket_list.py      # osmo bucket list
    ├── workflow_submit.py  # osmo workflow submit
    ├── workflow_list.py    # osmo workflow list
    ├── workflow_status.py  # osmo workflow describe
    ├── workflow_cancel.py  # osmo workflow cancel
    ├── workflow_logs.py    # osmo workflow logs
    ├── workflow_exec.py    # osmo workflow exec
    ├── task_list.py        # osmo task list
    ├── task_describe.py    # osmo task describe
    ├── data_upload.py      # osmo data upload
    ├── data_download.py    # osmo data download
    ├── data_list.py        # osmo data list
    ├── dataset_list.py     # osmo dataset list
    ├── dataset_describe.py # osmo dataset describe
    ├── app_list.py         # osmo app list
    ├── app_create.py       # osmo app create
    ├── cookbook_fetch.py   # OSMO cookbook fetcher
    ├── workflow_validate.py# local YAML lint
    └── workflow_render.py  # local Jinja render

Design tenets:

1. CLI is truth. Tools shell out to osmo, never re-implement its client SDK.
2. Thin wrappers. Each tool ≈ 50 lines, normalizing JSON into Strands ToolResult.
3. Graceful degradation. Missing osmo binary → exit 127 ToolResult, not crash.
4. No upstream forking. Use OSMO as-is; it's a separate repo on disk.

---

Workflow YAML 101 (cheat sheet)

workflow:
  name: my-pipeline
  resources:
    default: { cpu: 4, memory: 16Gi, storage: 40Gi, gpu: 1 }

  tasks:
  - name: train
    image: pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel
    command: ["python", "train.py"]
    args: ["--epochs", "{{epochs}}", "--num-gpu", "{{num_gpu}}"]
    resources: { gpu: "{{num_gpu}}" }

Submit with:

osmo workflow submit my-pipeline.yaml --pool h100-prod \
    --set num_gpu=4 --set epochs=10

Or:

osmo_workflow_submit("my-pipeline.yaml", pool="h100-prod",
                     set_vars={"num_gpu": 4, "epochs": 10})

For the full reference (including groups, inputs, outputs.dataset, checkpointing, and platform constraints) see the OSMO User Guide.

---

Development

git clone https://github.com/cagataycali/strands-osmo && cd strands-osmo
pip install -e ".[dev]"
pytest tests/                # smoke tests
ruff check .                  # lint
ruff format .                 # format

---

See Also

• OSMO - https://github.com/NVIDIA/OSMO
• OSMO docs - https://nvidia.github.io/OSMO/main/
• Strands Agents - https://strandsagents.com
• strands-cosmos - https://github.com/cagataycali/strands-cosmos

---

License

Apache 2.0 - same license as upstream OSMO.

OSMO is © NVIDIA Corporation. strands-osmo is an independent community project.