Every feature - levers9

Run code

feature	what it does	read
`@job`	run a Python function in a container: `.remote()` once, `.map()` one container per item	CPU job
`@endpoint`, `@asgi`, `@realtime`	a function, an ASGI app or a WebSocket app behind an HTTP URL with keep-warm and autoscaling	Endpoint
`@queue`	background work with retries, fed from Python or HTTP	Queue
`@cron`	a function on a cron expression (UTC)	Cron
`Box`	any long-running process with ports: servers, databases, notebooks	Box, shell and run
`Sandbox`	an isolated container you drive from Python: run code, manage files, expose ports, snapshot memory	Sandbox
`Blueprint`	Python version, pip and apt packages, shell commands, Docker-in-Docker; built once, content-addressed	Blueprint
`Volume`, `CloudBucket`	persistent files shared between containers; S3 or GCS buckets mounted directly	Volumes
secrets	API keys injected as environment variables	Secrets
GPUs	`gpu="L40S"` or a preference list; T4 to H200 and RTX PRO 6000	GPU
pipelines	chain functions on a volume: prep, train in parallel, evaluate, package	Pipeline, GPU pipeline
`VLLM`	serve an open-weights model behind an OpenAI-compatible URL	LLM serving
`MCPServer`	deploy a FastMCP server as a URL tools can call	MCP server
`Output`	save a file from inside a container and get a download URL	Parameters
`levers9 serve`	a temporary URL with hot reload while you develop	CLI
`levers9 shell`, `run`, `dev`	an interactive shell, a one-off container, a remote dev box	CLI

Watch and operate

feature	what it does	read
Apps	every function and deployment with its description (from the docstring), its calls in order, each labelled with its arguments, containers, logs and an execution timeline per call	Apps
Containers	what is running now, with live output and a stop button	Containers and workers
Workers and pools	the worker processes per node, pool, CPU, memory and GPU in use; workspace metrics	Containers and workers
Storage	a file browser per volume with upload, download, previews, and where each volume is mounted	Storage and secrets
Secrets	create from integration templates, rotate, delete	Storage and secrets
Notebooks	JupyterLab boxes with a native notebook editor, file browser, live CPU/memory/GPU meters, attached volumes and secrets	Notebooks
Endpoints	a curated open-weights model catalog with GPU fit, deployed with vLLM in one click	Endpoints
Workspaces	one per email domain or per team; invite members by email	Workspaces and environments
Environments	`main`, `staging`, a branch, a client: each a separate controlplane workspace; `levers9 login --env NAME`	Workspaces and environments
Sign-in	Google sign-in for the dashboard and the CLI, or a plain token when self-hosting	Sign-in
Empty states	every page that can be empty says what it is for with a drawn animation: a recording cut into labelled clips, a notebook cell running, files stacking on a volume, a secret masked into an environment variable, a request and its reply, a worker taking a container
Insights	interactive analytics over datasets, the Arena and tasks: build a chart or a number, click to filter everything, pin to a board, drill to rows, SQL and questions in words	Insights
Anonymization	one scan per dataset finds phones (written or said out loud), IDs, IBANs, cards, emails and names and writes a masked copy; a policy per dataset (Off, Flag, Mask) decides whether share links, published files, draft models and Insights read that copy; a review of a random sample of rows measures the scan (precision and recall per type, each with its range) and fixes the masking of every row reviewed	Personal data
Agent skills	a SKILL.md per feature for Claude Code, Cursor and Codex; `levers9 skills install`	Skills and MCP for AI agents
MCP servers	the docs as an MCP server, and the workspace as one: deployments, tasks, logs, GPU capacity, invoke (`levers9 mcp` or the dashboard API’s `/mcp`), and deploy from the local one	Skills and MCP for AI agents

Data Lab

feature	what it does	read
Datasets	import audio or text from volumes, S3, GCS, BigQuery, Hugging Face or URL lists; chunk audio into clips	Datasets
Decision gold sets	a human gold set for a decision model: reviewers answer typed questions about a lead, mark the events that caused each answer, say how it would change without an event and review the logged decision; exported as JSON lines with twin records, and the questions as a schema	Decision gold sets
Pre-labelling	optional drafts from Whisper, Azure, Google, ElevenLabs, Kokoro, vLLM, Anthropic or OpenAI, and from a Decision API for a decision dataset	Datasets
Annotation	a keyboard-driven editor for transcripts, roles, replies and clip quality, with statuses and filters	Annotation
Spreadsheet round-trip	export TSV, edit anywhere, import back	Annotation
Sharing	named links per annotator, an anyone-with-the-link option, viewer or annotator roles	Sharing
NDA gate	an agreement annotators sign before they see a row; versioned wording	Sharing

RL Arena

feature	what it does	read
Call logs	real calls in through a webhook, a dataset or paste; every reply scored on the rubric	Call logs
Scenarios	thousands of moments grouped automatically into a few dozen situations, on a map you can pan, zoom and select	Scenarios
Arena	simulated customers with personas and moods talk to a policy; fork any conversation at any turn	RL Arena
Compare	two replies for one moment; pick the better one and tick why	Train and evaluate
Train	preferences become a new policy version instantly; DPO with LoRA on a GPU for weights	Train and evaluate
Evaluate	two versions on the same customers and seeds: win rate, mean reward, per-dimension radar, regressions	Train and evaluate
Policies	versions with lineage; make one live	Train and evaluate
Culture pack	the market’s rubric, guardrails and personas; draft one from your playbooks	Culture pack
Tenants	one client per tenant, each with its own pack, webhook key and filters	Call logs