αlyxruntime

Use any other computer's CPU/GPU as your Jupyter runtime - like Colab's "connect to local runtime", but for any machine you control, from the command line or from Python.

The host talks out to Cloudflare through cloudflared, so there is no port forwarding, no public IP, and HTTPS is included - it works from behind home routers, NAT and most firewalls.

Install

pip install "alyxruntime[host]"   # on the machine that SHARES its GPU/CPU
pip install alyxruntime           # on the machine that CONNECTS (lighter)

Usage

On the host (the machine with the GPU):

alyxruntime stream --allowed_ip 203.0.113.7                # your client's public IP
alyxruntime stream --allowed_ip 203.0.113.7 10.0.0.0/24    # several IPs / CIDR ranges
alyxruntime stream --allowed_ip all                        # any IP - the token is the only gate

It starts a Cloudflare quick tunnel (if cloudflared is missing it offers to download the official binary, or pass --install-cloudflared) and prints a ready-to-paste command with a random access token:

alyxruntime connect 'https://random-words.trycloudflare.com/?token=...'

Not sure what your public IP is? Run alyxruntime connect from the client anyway: the host replies access denied for IP x.x.x.x. Ask the host to add it with --allowed_ip x.x.x.x.

On the client:

alyxruntime connect https://my-host:8765/?token=XXXX   # verifies, shows remote CPU/RAM/GPU, saves session
alyxruntime run train.py            # run a script   (also .ipynb, `-c "code"`, or stdin with `-`)
alyxruntime shell                   # interactive prompt; variables persist between commands
alyxruntime info                    # remote CPU / RAM / GPU
alyxruntime upload data.csv         # copy files to the host's working dir
alyxruntime download model.pt
alyxruntime restart                 # fresh kernel
alyxruntime disconnect --kill-kernel

Plots (plt.show()) are saved locally to ./alyx_outputs/. VS Code / JupyterLab also work: Select Kernel -> Existing Jupyter Server -> paste the same URL.

From Python

from alyxruntime import Runtime

rt = Runtime.connect("https://my-host:8765/?token=XXXX")
rt.run("import torch; print(torch.cuda.is_available())")
res = rt.run("2 + 2", echo=False);  print(res.results)     # ['4']
rt.upload("data.csv"); rt.download("out.pt")

Host side from Python: from alyxruntime import serve; serve(allowed_ip=["203.0.113.7"]).

Stable URL (named tunnel)

Quick tunnels get a new random URL every run. For a fixed hostname, create a tunnel in the Cloudflare dashboard (Zero Trust -> Networks -> Tunnels), point a public hostname at http://localhost:9000, then:

alyxruntime stream --tunnel-token <TOKEN> --public-url https://gpu.example.com --port 9000 \
                   --allowed_ip 203.0.113.7

stream options

flag	meaning
--allowed_ip IP [IP...]	public IPs / CIDR ranges (v4 or v6) allowed to connect, as seen by Cloudflare, or the word all (same as --allow-any). Required in tunnel mode.
--allow-any	accept any IP (the token is then your only protection)
--no-tunnel	skip Cloudflare; listen directly on --host:--port (you handle NAT; loopback is always allowed in this mode)
--install-cloudflared	download cloudflared automatically if it's missing
--tunnel-token / --public-url / --port	use a named tunnel with a stable hostname (see above)
--gpus 0,1	share only these GPU indices (none = CPU only). Run several streams with different --gpus to give each client its own GPU
--token	fixed access token instead of a random one
--root-dir	directory the client can see (default: current dir)
--tls-cert / --tls-key	HTTPS for --no-tunnel mode (not needed with a tunnel)
--public-host	hostname/IP to print in the connect URL with --no-tunnel

Security - please read

A runtime is remote code execution by design: whoever connects runs arbitrary code as the host user.

• Two independent gates: the IP allowlist and a random token.
• How the allowlist works with a tunnel: cloudflared reaches the host from 127.0.0.1, so the socket address is useless for filtering. The gate instead reads the CF-Connecting-IP header, which Cloudflare's edge sets itself, and checks every request against --allowed_ip. Requests with no such header (anything not coming through the tunnel) are refused, loopback gets no special pass, and each request is forced to Connection: close so one vetted connection can't be reused by someone else.
• The quick-tunnel URL is public: anyone who learns it can reach the gate (and be refused). Treat the token like a password, and don't paste the connect line in public places.
• Traffic passes through Cloudflare's network (TLS to Cloudflare, then Cloudflare -> your host over the tunnel); Cloudflare can technically see it. If that's unacceptable, use --no-tunnel with your own VPN.
• IP allowlisting trusts that Cloudflare reports the true client IP. It won't protect against someone sharing your NAT/VPN exit IP, and mobile/CGNAT or rotating-IP connections will need a CIDR range.
• Local users on the host can reach the gate on its loopback port and are trusted like the host user.
• Don't run stream as root unless you mean it. The client's saved token lives in ~/.alyxruntime/session.json (0600).
• Cloudflare's free quick tunnels are meant for development: no uptime guarantee, and you should read Cloudflare's terms before using them for anything heavy or long-lived (use a named tunnel for that).

Web UI (Colab-style notebook)

alyxruntime connect <url>   # once (or use the Connect button in the UI)
alyxruntime ui              # opens http://127.0.0.1:8899 - notebook cells run on the remote GPU/CPU

Code/text cells, Shift+Enter / Ctrl+Enter, streaming output and plots, run all / interrupt / restart, remote file browser with upload/download, open/download .ipynb, dark mode. Notebooks autosave in the browser. The UI only listens on 127.0.0.1 and requires a per-launch secret, so other websites can't drive your runtime.