urun CLI

Deploy Python apps to urun from your terminal.

Install

uv tool install urun-cli
# or
pip install urun-cli

The package installs the urun command:

urun --version

For one-off uvx usage:

uvx --from urun-cli urun --version
# or the package-matching command alias
uvx urun-cli --version

Quick start

Save an org-scoped deploy API key locally with urun login:

urun login

If URUN_API_KEY is not already set, urun login opens the urun console API keys page, prompts you to paste the generated urun_sk_... key, verifies it with the urun API, and stores credentials for later commands.

If you already have a key, you can pass it directly:

urun login --api-key urun_sk_<secret>

For CI or one-off commands, you can still use the environment variable:

export URUN_API_KEY=urun_sk_<secret>

Create app.py:

import urun
from urun import App

app = App("hello-h100")


@app.function(gpus="h100:1")
def hello(ctx: urun.Context):
    print(f"running on {ctx.device}")
    return {"device": str(ctx.device)}

Run it:

urun run app.py

In this release, urun run uses the same deploy pipeline as urun deploy. deploy remains available as the lower-level command while the full deploy/run/monitor workflow is being built.

Launch an official browser demo

For official urun-examples apps, the shortest browser stream path is:

cd ~/workspace/urun-examples
urun deploy prompt_canary/backend/app.py --name prompt-canary --timeout 900
urun demo prompt-canary --examples-root ~/workspace/urun-examples

urun demo reads urun-demo-manifest.json from the trusted examples checkout you pass with --examples-root (or URUN_EXAMPLES_ROOT) and supports manifest entries with demo.supported=true. It uses your stored API key server-side to request a short-lived browser JWT, sets the frontend's manifest-declared NEXT_PUBLIC_* environment variables, starts the local Next.js frontend, and opens it in your browser.

Use the manifest slug as the positional argument. If you deployed an official example under a custom app name, pass it explicitly:

urun demo matrix-game-3 --app test-matrix-game

Override the examples checkout with --examples-root or URUN_EXAMPLES_ROOT; override the session API with --session-url or URUN_SESSION_BASE_URL.

Serve a model from the catalog

urun serve deploys a model straight from the urun model catalog — no app code required. It is urun deploy with a serve app templated from the resolved catalog row (engine, HF repo, GPU, engine args).

List the enumerable model matrix (models, variants, the GPU each fits, engine shape, and whether the placement is dev-testable or prod-only):

urun serve catalog
urun serve catalog --json

Serve a model. With no variant the model's default (first) variant is used; with no --gpu the variant's first placement is used:

urun serve qwen-coder                    # default variant + first placement
urun serve glm-5.2:UD-IQ2_M              # explicit <id>:<variant>
urun serve qwen-coder-480b:fp8 --gpu b200:8

The catalog is anon-readable reference data published by a urun-infra migration. Reads go over PostgREST on the same control-plane host as the API URL; supply the public Supabase anon key via --anon-key or URUN_CATALOG_ANON_KEY (URUN_SUPABASE_ANON_KEY is also honored). Override the PostgREST base with --catalog-url / URUN_CATALOG_URL if needed.

By default urun serve <id> renders a templated serve app and deploys it. If you maintain your own _serve app, point --entrypoint at it; the CLI deploys that app with URUN_SERVE_CONFIG and URUN_SERVE_CATALOG_ROW set to the resolved row instead of rendering one.

Inspect apps

List every app deployed in your org and its current status:

urun list apps

Sample output:

APP                    FUNCTION        COMPUTE  STATUS         RELEASE       DETAIL
causal-forcing-stream  generate_video  h100:1   ready          fa61f31b0961  0/1 GPU units in use
queued-app             warmup          a10:1    provisioning   000000000000  building

The STATUS column is one of provisioning, ready, pending, paused, or failed. It is derived from three backend signals reporting on sequential lifecycle phases:

The DETAIL column carries the disambiguating signal (raw build state, error message, ready_reason, or in-use GPU counts). Pass --json for the raw payload.

This command is experimental and requires the server-side GET /apps endpoint, which is in development.

Inspect sessions

List live and historical sessions in your org. Newest sessions appear at the bottom of the table so the command works well with tail:

urun list sessions
urun list sessions | tail -20
urun list sessions --state failed
urun list sessions --limit 500

Sample output:

ID            APP                FUNCTION        SHAPE    STARTED               DURATION  STATE       DETAIL
2cc8a91f4b3d  helios             world_gen       h100:4   2026-06-02 10:55 UTC       44s  failed      no_capacity
3f0017daee01  helios             world_gen       h100:1   2026-06-02 11:08 UTC    18m43s  completed   client_disconnect
4a1c886e2d0a  causal-forcing-…   generate_video  h100:1   2026-06-02 14:21 UTC     3m12s  live        -

The STATE column maps the raw backend status to a user-friendly label:

DURATION is computed from allocated_at to closed_at for terminal sessions, or allocated_at to now for live ones. DETAIL carries close_reason when present. Pass --json for the raw payload (full IDs, ISO timestamps, all fields).

Pass --limit to control how many rows are fetched (default 100).

This command is experimental and requires the server-side GET /sessions endpoint, which is in development.

Inspect active compute

List the compute slices your org currently has provisioned:

urun list compute

Sample output:

APP                    FUNCTION        SHAPE   INSTANCES  GPU UNITS  SESSIONS  AGE
causal-forcing-stream  generate_video  h100:1  1/2        1/2        1         12s
helios                 world_gen       h100:4  0/1        0/4        0         3m

Each row is one actively provisioned (app, function, compute_shape) slice. INSTANCES and GPU UNITS show <allocated>/<provisioned> — a row with 0/1 is an idle warm runtime with no active sessions on it. SESSIONS is the live session count. AGE is how stale the capacity snapshot is; very old ages may indicate the runtime is no longer reporting.

Slices with no provisioned capacity are omitted, so this command answers "what is running right now". For the full deployment catalogue (including paused / failed / unprovisioned apps) use urun list apps; for historical or in-flight sessions use urun list sessions.

Pass --limit to control how many rows are fetched (default 100).

This command is experimental and requires the server-side GET /compute endpoint, which is in development.

Manage apps

Manage the lifecycle of a single deployed app. The app is addressed by its slug (the name shown under APP in urun list apps); every operation is org-scoped via your API key.

Show detailed status for one app (the single-app complement to list apps):

urun app status lingbot

            App: lingbot
           Name: LingBot
    Environment: prod
     App status: active
     Deployment: active
Desired replicas: 2
       Function: handle_lingbot_runtime
        Compute: b200:4
            GPU: 4 x b200
        Release: 1c6d6287abcd
  Live sessions: 1

Scale an app's runtime replica count (the backend's scaling knob; the control plane turns it into the runtime StatefulSet replica count):

urun app scale lingbot --replicas 3
urun app scale lingbot --replicas 0   # drain to zero without retiring

GPU count and compute shape are fixed at deploy time per release (set via @app.function), so scale intentionally exposes only --replicas.

Disable an app so the control plane stops running it (drives the deployment to paused and the app to disabled, so the materializer stops recreating its runtime). This is the clean, reversible, API-driven alternative to a manual database edit:

urun app disable lingbot-handle        # prompts for confirmation
urun app disable lingbot-handle --yes  # skip confirmation

Enable a disabled app and bring it back online:

urun app enable lingbot-handle

Stop a queued, starting, or live session:

urun stop session sess_123 --yes

All app subcommands accept --environment (default prod) and --json. These commands are experimental and require the server-side app lifecycle endpoint, which is in development.

Scratch instances (rung-4 verification, dev-only)

The verification ladder for platform changes is: unit tests -> CPU harness -> kind hop -> scratch pod -> full deploy. Rungs 1-3 are one command each; urun scratch makes rung 4 one command too — a real GPU pod running the real render, without the 30-90 min deploy loop and without touching the live app:

# Reproduce a broken app in isolation with debug env, candidate wheel overlaid:
urun scratch gemma-voice dg-brain --shape rtx6000:1 \
  --env VLLM_LOGGING_LEVEL=DEBUG --env CUDA_LAUNCH_BLOCKING=1 \
  --env PYTHONFAULTHANDLER=1 --name dg-brain-noble-repro

urun scratch ls                      # list scratch instances + expiry
urun scratch rm dg-brain-noble-repro # tear down
urun scratch rm --expired            # reap everything past its TTL

What it does:

• Clones the app's rendered StatefulSet into an isolated instance with every platform label detached: the materializer never adopts it, the capacity reconciler never counts it, and no session is ever routed to it. The clone boots even when the source app is crashlooping or scaled to 0.
• Overlays candidate artifacts: --wheel <req|url|path-on-storage> installs --no-deps into a hardlink COPY of the deps venv (the shared venv is never modified); --env KEY=VAL adds debug vars; hot reload is always pinned off so the instance stays on the currently-rendered release (--release <hash> asserts which one that is).
• Takes GPUs explicitly: --shape is required and must match the source render; the target karpenter nodepool is checked against its GPU limit and the command refuses at capacity unless --allow-contention. Scratch never silently steals demo GPUs.
• Never leaks: every instance carries a TTL annotation (default 4h, --ttl 6h/90m) consumed by the dev scratch reaper; urun scratch rm tears down sooner. Boot/self-test logs stream to your terminal (Ctrl-C detaches without tearing down).

Dev-only (v1): talks to the dev cluster via your kubeconfig (--context, default dev-usw2) and refuses prod contexts.

What gets deployed

urun deploy creates a source manifest from your Python entrypoint:

Dependencies are declared in your urun app code. Project-level files such as pyproject.toml and requirements.txt are not uploaded as dependency declarations by the CLI.

Generated/cache content such as .git, dotfiles, __pycache__, and .pyc files is excluded. Add .urunignore to exclude additional paths.

Non-Python assets such as templates, static files, and data files are not auto-included yet.

Common options

Shared by run and deploy:

Troubleshooting

Development

Contributing and test instructions are in CONTRIBUTING.md.

License

MIT.

Development environment

This repo has a Nix/direnv/devcontainer baseline:

direnv allow
just sync
just check

Use VS Code Dev Containers to open the repository with the same toolchain in a container. Copy devcontainer.env.example to .devcontainer.env if you need to pass local git identity or other non-secret development settings into the container.