Skip to content
Conduit

skill

neko-computer

small computers for ai agents

superhq-ai/neko-computercommunitylistedunreviewed

Set up

Skills are installed by placing their folder where your agent loads skills.

The download comes from github.com/superhq-ai/neko-computer itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 26, 2026

Expose a sandbox or a local port to the public web with one command, at <name>.neko.computer.

Install

macOS and Linux, both aarch64:

curl -fsSL https://install.neko.computer | sh

Or with Homebrew: brew install superhq-ai/tap/neko. Update later with neko upgrade.

Usage

neko login                              # sign in via the SuperHQ device grant (app.superhq.ai)
neko tunnel 3000                        # tunnel a local 127.0.0.1:3000 port (no sandbox)
neko tunnel 3000 --workspace acme       # own the public tunnel in a SuperHQ workspace

neko run --port 3000 -- npm run dev     # reach the sandbox from your machine, no account
neko run --tunnel 3000 -- npm run dev   # anonymous, ephemeral sandbox, public URL
neko run --tunnel 3000 -w acme -- npm run dev
neko run alice --tunnel 3000 -- npm run dev            # a persistent named computer
neko run base --allow-net --checkpoint clean -- sh -c 'apt-get install -y python3'   # build a reusable checkpoint
neko run web --from base@clean --tunnel 3000 -- python3 -m http.server 3000

Managing computers and checkpoints:

neko clone base@clean staging   # a computer from a ref (shared node) or an image file
neko ls                         # list computers
neko history alice              # the checkpoint history of a computer
neko rm alice                   # remove a computer and reclaim its orphaned checkpoints
neko gc                         # sweep checkpoints orphaned by crashes or interrupted ops

--port forwards a sandbox port to 127.0.0.1 on your machine, which is all local: no account, no sign-in, nothing hosted. Use --port 3000:8000 for a different host port. --tunnel is the same plumbing pointed at the public internet instead, which needs SuperHQ Pro.

A computer is a named, versioned sandbox: a branch ref over a shared tree of checkpoints (immutable disk snapshots). neko run NAME resumes from the computer head; --from branches from a checkpoint; --checkpoint snapshots on the way down. Opening a tunnel requires neko login.

Sandboxes

neko run boots an isolated Linux microVM, mounts the current directory read-only at /workspace, and runs the command there. argv is passed directly, with no shell, so pipes and && need an explicit sh -c. The base image is a minimal Debian: install what the command needs inline, or bake it into a checkpoint. The image downloads on first run.

What the sandbox may touch

A sandbox starts closed, and each door is opened by a flag. Both defaults are printed at boot, so a command that trips over one has the reason in view.

neko run --allow-net -- sh -c 'apt-get update && apt-get install -y python3'
neko run --allow-host '*.npmjs.org' -- npm install    # that host and no other
neko run --write -- npm run build                     # keep the output on the host
neko run --mount ~/data:/data --mount .:/src:rw -- ./build.sh
  • Networking is off. The guest boots with no network device at all, so nothing reaches the internet and nothing leaves. --allow-net opens it; --allow-host PATTERN (repeatable, implies --allow-net) opens only the hosts you name, through a host-side proxy. Tunnels and --port are unaffected either way: they reach the guest over vsock, not over its network.
  • The current directory is mounted read-only at /workspace. This is not a wall the command hits: the guest lays an overlay over the share, so npm install and build output succeed and land in a scratch layer that goes away with the sandbox, leaving your tree untouched. --write shares it read-write instead, so writes are real. --mount HOST:GUEST[:ro|:rw] mounts elsewhere, is repeatable, and is read-only unless it ends in :rw; --workdir sets where the command runs, defaulting to the first mount's guest path.

Installing at every boot is both slow and a reason to keep egress open. Bake the runtime into a checkpoint once with --allow-net, then boot from it offline:

neko run base --allow-net --checkpoint node -- sh -c 'apt-get update && apt-get install -y nodejs npm'
neko run web --from base@node --tunnel 3000 -- npm start

Web terminal

neko run --term puts a terminal into the sandbox on the web: it opens a private tunnel of its own and serves the terminal at https://<name>.neko.computer/__neko/term (the bare URL redirects there). Only members of your workspace can open it (--private SLUG to name one), and each visit gets its own login shell. Add --tunnel PORT to expose an app port on the same tunnel alongside the terminal.

neko run --term                         # a computer in the browser
neko run alice --term                   # a persistent one
neko run --term --tunnel 3000 -- npm run dev   # app plus terminal

Agent skill

neko ships as an agent skill so coding agents (Claude Code, Cursor, Copilot, and more) can open a tunnel whenever they need a public URL for a sandbox or a local port.

# Install with the skills CLI
npx skills add superhq-ai/neko-computer

# Or copy it into your project manually
cp -r skills/neko .claude/skills/neko

The skill lives at skills/neko/SKILL.md.

Self-hosting

The edge worker is a library: deploy your own tunnel domain and point the CLI at it. See docs/self-hosting.md.

Layout

  • crates/neko the Rust CLI
  • edge the Cloudflare tunnel Worker and Durable Object
  • skills/neko the agent skill
  • docs the tunnel protocol, self-hosting, and development notes

The wire protocol is specified in docs/specs/tunnel-protocol.md.

License

Apache-2.0. See LICENSE.

↑↓ navigate openesc close