autodl-cli

Unofficial. This project is not affiliated with, endorsed by, or sponsored by AutoDL. "AutoDL" is used only to identify the platform this tool talks to.

Manage AutoDL GPU instances from the command line — and let your coding agent do it too.

Built on AutoDL's official open API, so the developer token stays valid indefinitely and nothing breaks when the web console is redesigned.

简体中文

Why

When an AI agent needs a GPU box mid-task, it has no way to get one: AutoDL's instances are created by clicking through a web console. This gives the agent three ways in, all backed by the same core:

Every command speaks --json with a stable schema and a documented exit code, so an agent can branch on the result without parsing prose.

Related MCP server: lab-ops

Install

npm install -g @minatoaqukin/autodl-cli   # then: autodl <command>
npx @minatoaqukin/autodl-cli <command>    # or without installing

Requires Node.js 22+.

The dashboard

autodl tui     # or just `autodl` in an interactive terminal

On a machine with no token yet it opens on a login screen offering two things — configure a token, or quit. Paste the token, it is verified against the API before being saved, and you land on the dashboard. Once configured, autodl goes straight in.

A live table of your instances: status, GPU, region, how long each has been powered on and roughly what that has cost, and how much TTL is left. Keys: ↑↓ move, Enter detail, s start, x stop, c copy the SSH command to the clipboard, Ctrl+D release, g GPU stock, n new instance, r refresh, q quit. Release sits on Ctrl+D rather than a bare key because it wipes the instance permanently.

It fills the terminal and shows your account id and balance in the header. Copying with c puts only the SSH command on the clipboard — never the root password, which any process could then read. Where no clipboard helper exists (SSH sessions, containers) it falls back to OSC 52 and says so, since the terminal never confirms.

It runs on the terminal's alternate screen, so it owns a fixed canvas instead of scrolling below whatever was already there, and quitting restores your prompt and scrollback untouched.

It exists because AutoDL bills on power state: the expensive mistake is not a wrong command, it's an instance nobody remembered to stop. Leaving this open makes that visible.

Two deliberate honesty constraints. A rate is only knowable from a running instance's snapshot, so a stopped instance shows elapsed time and no money — inventing a number would be worse than showing none. And while a rate is still loading the total says so rather than quietly under-reporting.

The TUI never runs in a pipe, in CI, or under --json: it exits with code 2 and an explanation instead of taking over a terminal that isn't there. A bare autodl outside an interactive terminal still prints help exactly as before.

Setup

The official API needs a developer token from an identity-verified account (个人或企业实名认证). Get it from the AutoDL console → 设置 → 开发者 Token.

autodl login              # verifies the token, then saves it with 0600 permissions
autodl account            # balance, vouchers, lifetime spend

Token precedence: --token › AUTODL_TOKEN › ~/.config/autodl-cli/config.json.

Quick start

# Rent a GPU that shuts itself off after two hours, and wait until it's ready
autodl create --gpu 4090 --ttl 2h --wait

# Work with it
autodl ls
autodl ssh pro-76419909953e                    # interactive login
autodl exec pro-76419909953e "nvidia-smi"      # one-off command, remote exit code
autodl push pro-76419909953e ./src /root/work  # SFTP upload
autodl pull pro-76419909953e /root/work/out .  # SFTP download

# Stop paying
autodl stop pro-76419909953e
autodl rm pro-76419909953e --yes               # irreversible: wipes all data

Or do the whole thing in one verb:

autodl run "python train.py" \
  --gpu 4090 --sync ./ --pull /root/autodl-cli/checkpoints --ttl 4h

That creates an instance, waits for it, uploads your code, streams the command's output, downloads the results, and powers the instance off — including on Ctrl-C.

Use it from an agent

Claude Code

claude mcp add autodl -- npx -y @minatoaqukin/autodl-cli mcp

Cursor / Cline / any MCP client

{
  "mcpServers": {
    "autodl": {
      "command": "npx",
      "args": ["-y", "@minatoaqukin/autodl-cli", "mcp"],
      "env": { "AUTODL_TOKEN": "your-token" }
    }
  }
}

Tools exposed: autodl_account_info, autodl_list_instances, autodl_get_instance, autodl_create_instance, autodl_power_on, autodl_power_off, autodl_release_instance, autodl_exec, autodl_upload, autodl_download, autodl_run, autodl_list_gpu_specs, autodl_list_images, autodl_save_image, autodl_sweep_expired. Plus an autodl://instances resource.

MCP defaults are deliberately stricter than the CLI's, because nobody is watching: a 2-hour TTL is applied unless you ask for longer, releasing requires an explicit confirm: true, and passwords come back redacted unless requested.

Shell / CI

autodl ls --json | jq -r '.data[] | select(.status=="running") | .uuid'

Deploying a git project

# Rent a box, clone, auto-install dependencies, run it, then stop the instance
autodl deploy owner/repo --gpu 4090 --start "python train.py" --ttl 4h

# Long-running service: background it and keep the instance up
autodl deploy owner/repo --gpu 4090 --start "python app.py" --detach

# Come back later — powers the same box on and `git pull`s, no rebuild
autodl deploy owner/repo --instance pro-76419909953e --start "python train.py"

deploy differs from run in one deliberate way: it stops the instance at the end instead of releasing it. A stopped instance keeps its disks, so the next deploy reuses the environment you already built. --on-finish release opts out.

Code lands in /root/autodl-tmp/<repo> — the data disk. AutoDL's system disk is a fixed 30GB that also gets packed into any saved image; the data disk is separate, faster and expandable. The trade-off worth knowing: data-disk contents are not included when you save an image, so put the environment on the system disk and the code here.

Dependencies are auto-detected in this order, first hit wins — environment.yml → requirements.txt → pyproject.toml → package-lock.json/package.json. Override with --setup "<cmd>", or skip with --no-setup.

Remote commands run through a login shell. AutoDL images keep python, pip and conda in /root/miniconda3/bin, which only reaches PATH via the login profile — a plain non-interactive ssh host "pip install ..." exits 127. This applies to autodl exec too, so it behaves the way it does when you autodl ssh in by hand.

Cloning from GitHub or HuggingFace automatically enables AutoDL's academic proxy (source /etc/network_turbo). Gitee is domestic and skips it. --no-accel disables it. AutoDL notes the proxy is "for academic use, with no stability guarantee".

Private repos: --git-token, or GIT_TOKEN / GITHUB_TOKEN in the environment. The token never reaches a log line, an error message, --json output, or the checkout's stored git remote.

Checking GPU stock

autodl stock --gpu 4090        # where are the free cards
autodl stock                   # everything, everywhere

Read this table carefully — the numbers are less authoritative than they look. They come from AutoDL's elastic-deployment stock endpoint, the only capacity API that exists, and it does not track Pro instance availability. Measured on 2026-08-23: it reported 140 idle RTX 4090D in westDC3 while creating a Pro instance there answered "暂无库存" — and the identical request with no region constraint succeeded, landing in beijingDC2.

Two consequences, both baked into the tool:

• Creating an instance never narrows regions on its own. Omitting data_center_list gives AutoDL the widest choice, which empirically succeeds most often.

• Only two regions accept a Pro instance at all: westDC3 (西北B区) and beijingDC2 (北京B区). The other nine in the stock table are elastic-deployment only; passing one to --region is rejected up front rather than failing later with AutoDL's opaque "请求参数错误". The 可建Pro column marks which is which.

The cost guard

AutoDL bills purely on power state. An instance that finished training an hour ago costs exactly as much as one at 100% utilisation. This is the single easiest way for an unattended agent to waste real money, so the protection is built in rather than optional.

Three layers:

1. Inside the instance. --ttl 2h arms a detached sleep && shutdown on the box itself via start_command. It fires even if this CLI is killed, your laptop sleeps, or the network dies. This is the layer that actually protects your wallet.

2. A local ledger. Every command opportunistically sweeps instances past their TTL and powers them off. Catches the cases layer 1 can't — a start_command that silently failed, or a manual power-on with no fresh timer.

3. Idle detection. autodl guard idle <id> samples GPU utilisation over SSH and shuts down after a sustained lull.

autodl guard ttl pro-xxx 2h     # arm/re-arm on a running instance
autodl guard cancel pro-xxx     # disarm
autodl guard list               # what this machine is tracking
autodl guard sweep              # reclaim everything past its TTL now
autodl guard idle pro-xxx --threshold 5 --samples 6 --interval 1m

A balance gate also refuses to create an instance when your wallet is below --min-balance (default ¥5). AutoDL doesn't reclaim instances the moment the balance hits zero — it keeps them to protect your data — so a low balance turns into a stuck, unusable instance rather than a clean failure.

The agent contract

Stable across minor versions. Breaking changes require a major.

stdout in --json mode is pure JSON. Progress, prompts and warnings all go to stderr, so autodl ... --json | jq is always safe.

// success
{ "ok": true, "data": { /* ... */ } }

// failure
{ "ok": false, "error": { "code": "NO_STOCK", "message": "…", "hint": "…", "requestId": "…" } }

autodl exec and autodl run instead exit with the remote command's exit code, so autodl exec box "make test" && deploy behaves the way you'd expect.

Commands

Global flags: --json, --yes, --token, --base-url, --lang zh|en, --verbose, --no-color, --no-sweep.

push and pull skip .git, node_modules, __pycache__, .venv and friends, then apply .autodlignore if present, falling back to .gitignore.

SDK

import {
  AutoDLClient,
  createInstance,
  execCommand,
  powerOffInstance,
  waitForRunning,
} from "@minatoaqukin/autodl-cli";

const client = new AutoDLClient({ token: process.env.AUTODL_TOKEN! });

const uuid = await createInstance(client, {
  gpuSpec: "v-48g",
  gpuNum: 1,
  imageUuid: "base-image-l2t43iu6uk",
  cudaFrom: 118,
});

await waitForRunning(client, uuid);
const { stdout } = await execCommand(client, uuid, "nvidia-smi");
console.log(stdout);
await powerOffInstance(client, uuid);

Everything re-exported from the package root is public API. Prices arrive as yuan (number), timestamps as ISO strings, and Go's sql.NullTime shape is flattened to string | null.

What the official API cannot do

These are AutoDL's limits, not this tool's. Knowing them up front saves a lot of confusion:

• Pay-as-you-go only. No daily/weekly/monthly plans and no renewal endpoint.

• Pro instances only. The seven specs in autodl gpus — the cheaper standard instances aren't reachable through the open API.

• No usable stock query for Pro. The one capacity endpoint reports elastic-deployment stock, which demonstrably does not match Pro availability (see above). Creation is effectively a blind attempt; no capacity means exit code 6 and another spec to try.

• Only two regions accept a Pro instance: westDC3 and beijingDC2.

• No CPU-only boot yet. AutoDL's own wording is deliberately provisional: payload is documented as "gpu：有卡开机, 暂不支持API以无卡模式开机" — not yet supported, rather than never. Confirmed on a live instance 2026-08-24: cpu, no_gpu, nogpu, cpu_only, cpu-only and none all return ServerError | 不支持的启动模式, and an empty payload is accepted but boots with the GPU attached (start_mode: "gpu"). Use the web console for the ¥0.1/hr 无卡模式 in the meantime; this tool will expose it once the API does.

• Identity verification required before the API will respond at all.

• Missing operations: rename, scheduled shutdown, resizing, migration, system reset.

• SSH credentials can change on any power cycle — port and root password. AutoDL may reschedule the instance onto a different machine. It doesn't always happen (a real stop/start was observed keeping both identical), which is precisely what makes caching dangerous: a stale value works often enough to hide the bug until it doesn't. This tool re-reads them on every connection, so you never have to think about it.

• running does not mean sshd is ready. A freshly created instance reports running before it accepts connections. Connection attempts here are spaced out rather than fired back to back.

• A non-interactive SSH session has almost no PATH. No python, pip or conda — they live in /root/miniconda3/bin and arrive only through the login profile. Every remote command here runs under bash -lc for that reason.

• Releasing requires a completed shutdown, and a second power_off on an instance that is already stopping is an error. Both are handled internally.

Also worth knowing: an instance left shut down for 15 consecutive days is released and its data wiped.

Verified against the live API on 2026-08-23: full lifecycle (create → SSH exec → SFTP round trip → stop → start → exec again → release) on a 4090D, total cost ¥0.10.

The GPU spec, region and base-image tables are baked in because the API exposes no catalogue endpoint. If AutoDL changes them, please open an issue.

Development

npm install
npm run build
npm test            # 325 tests, no network access, no cost
npm run lint
npm run typecheck

Real end-to-end tests rent an actual GPU and cost actual money, so they're opt-in:

AUTODL_E2E=1 AUTODL_TOKEN=<token> npm run test:e2e

They always power the instance down in an afterAll, even on failure. Add AUTODL_E2E_RELEASE=1 to release it too.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Especially valuable: corrections to the static catalogue, and real API error codes we haven't mapped yet (AutoDL doesn't document them).

License

MIT