husk

Husk gives your AI chat named local workspaces. Save public web sources, ask your AI for a cited brief, comparison, or action list, then review and download the result. Reopen the same workspace in your next chat. No Husk account, Docker, terminal, or additional model API key is needed for this starter flow.

For coding and automation, Husk also provides optional computer tools: a shell, files, and a browser. Existing MCP and CLI computer workflows remain available.

Start with a workspace

Download husk-0.2.0.mcpb from the public preview. Install it in Claude Desktop through Settings → Extensions → Advanced settings → Install Extension. Claude Desktop supplies Node.js. This candidate is unsigned; organization policies may restrict custom extensions.

Ask Claude: “Create a Husk workspace called My first brief and open its viewer.” Add public page URLs, choose a starter prompt, paste it back into Claude, then return to review and download the saved result. Keep Claude running while the viewer is open.

The workspace guide explains the full flow, privacy, recovery, and limits. To build the installer as a contributor:

npm ci
npm run build
npm run bundle:mcpb
npm run smoke:starter -- --bundle

The artifact and checksum are written to build/. The smoke test extracts that archive and checks restart persistence, access control, assets, downloads, and ZIP export.

Open interactively — pan, zoom, trace relationships · SVG · all 16 diagrams

Related MCP server: linux-harness-mcp

Computer tools for developers

npx @husk-ai/cli@~0.2 setup

setup asks where your chats' computers should run (Docker, Podman, this machine, SSH or Fly, with the reason any of them is unavailable and its fix), adds Husk to the chat apps it finds (Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Zed, Codex, Gemini CLI), and starts the one Husk every chat goes through, printing its console link. It backs up each app's config file before changing it and touches only its own husk entry. --remove takes it out again. The console's Connect page does the same. What each app supports, and how Husk was checked against it, is in docs/clients.md.

Coming from 0.1.4? Run setup again, or Update on the Connect page. Each chat app's husk entry moves to this release and keeps the flags and env you gave it, Claude Code's per-project entries included (claude mcp add writes those by default). Your computers, the named sessions bound to them, their files and your bots carry over as they are.

claude.ai and ChatGPT run in someone else's data centre, so they reach Husk through a tunnel you run, or a Fly app of your own: husk serve --remote --public-url https://<name>. It is for you alone (a claim code once, then your passkey), and every command, file change, click, typed text, upload and script waits for your passkey; reading pages and files does not. Who can read what, the tunnels and the off switch are in the remote access guide.

Once the next release is on npm, you can let the chat do it. Paste this into Claude Code, Codex, Gemini CLI, Cursor or any chat that can run a command (the Connect page makes one for a single app). Until then the npx line in it fails, because npm has no 0.2 yet: use the checkout command above instead.

Set up Husk for me, so you get a computer of your own for tasks: a terminal and a real browser, kept apart from my files.

1. Run this in a terminal:
   npx -y @husk-ai/cli@~0.2 setup --computer auto --apps all --yes
   It picks the safest computer this machine can run (a Docker container if Docker is installed) and adds Husk to my chat apps. If you can't run terminal commands, show me that line and I'll run it.
2. Tell me to restart my chat apps so the Husk tools load.
3. Tell me that if the Husk desktop app is installed, it shows your computer in a small window and I can take over any time.

After the restart: call computer_info, open https://example.com in the browser, and tell me what you see.

A desktop app (apps/desktop) carries the same Husk with its own Node and keeps each chat's computer one click away: a very small 3D computer for each in the bottom-right corner, which names its chat app when you hover it. Click one and a compact Computers window opens, its tabs your computers (each with its chat app's logo), the browser live on a white screen, Take over a click away. Minimize it and it goes to a notch at the top of the screen (around the real notch on a Mac). It sits in the tray to tell you when a chat is waiting for you. Each release attaches its installers for Windows, macOS and Linux, signed and notarized once the release has the signing keys and unsigned until then; see apps/desktop.

To add Husk to one app by hand instead:

claude mcp add husk -- npx -y @husk-ai/mcp

That is the whole setup. Nothing is created until the first tool call, so an installed husk that nobody uses costs nothing. Works with Cursor, Zed, or anything else that speaks MCP.

In Claude Code you can install it as a plugin instead, which adds a husk skill that teaches the model when to use which tool and when to hand the computer to you:

/plugin install husk --marketplace Hotragn/husk

The default computer profile exposes eight workspace tools plus 26 computer tools. Eight are for the machine — shell, read_file, write_file, edit_file, list_dir, expose_port, browse, computer_info — fifteen browser_* tools that drive a real Chromium inside that same machine, addressed by accessibility ref rather than pixel coordinates, hand_off / hand_off_done, which give you the computer for a login, a 2FA code or a CAPTCHA and give it back only when you say you're done, and computer_open_ui, which gives the chat a link to show you the computer. Apps that render MCP Apps cards also get the card's two tools. Apps that give up on a long tool call get job_wait: Husk answers such an app before it gives up, saying the call is still running, and job_wait collects the result.

The first tool result says what kind of machine it got. A model that thinks it is sandboxed when it is not makes worse decisions than one that knows.

Every chat on the machine goes through one shared Husk, started on the first tool call, so husk serve's console shows each chat's computer, its live browser and what it did last, and takes over a handoff from any of them. HUSK_DAEMON=off keeps a chat's tools in its own process.

The CLI

npx @husk-ai/cli onboard                 # guided setup, five steps, start here
npx @husk-ai/cli doctor                  # or go straight to what this machine can offer
npx @husk-ai/cli up dev --flavor python  # bring up a Linux computer
npx @husk-ai/cli exec dev -- 'uname -sr && python3 -V'
npx @husk-ai/cli rm dev                  # tear it down

onboard names your provider and how isolated it really is, creates a computer and destroys it in front of you, points at the shortest route to a model if you have none, and prints the MCP line for your editor. It never asks for an API key — husk reads credentials from the environment and does not store them, so it prints the export line and re-checks.

On Windows cmd.exe, use double quotes -- it does not strip single ones, so the quotes would reach the container as part of the command:

npx @husk-ai/cli exec dev -- "uname -sr && python3 -V"

--flavor python is there because the default base flavour is debian:bookworm-slim, which has no python3 -- and a quickstart whose third line prints python3: not found is not a quickstart. On a laptop with Docker running, that sequence prints Linux 6.18.33.2-microsoft-standard-WSL2 and Python 3.12.14; with Docker stopped and WSL2 installed it picks local and prints the same kernel with the distro's own Python 3.14.4. doctor reports every provider it probed, which one it would pick, and why the others were skipped.

Turn a chat into a bot

A conversation that worked once becomes a file you can run again.

husk import                          # lists transcripts it found
husk import --pick 1                 # imports one, prints its id
husk distill <id> --out triage.yaml  # conversation → agent spec
husk run triage.yaml "check the build"

Importers read Claude Code JSONL, Codex rollouts, ChatGPT, Cursor, Gemini and markdown into one Transcript. In the console, Husks › Make a bot from a chat does the same from a chat it finds on this computer, an export, or pasted text. A branched Claude Code thread is rebuilt along its longest parentUuid chain, the latest leaf breaking a tie, so the dead ends you backed out of are left out. The distiller runs with no API key at all, and the model-backed mode falls back to that free path rather than failing. Secrets are stripped before the file is written, not after.

What comes out is YAML you can read and diff. This is husk distill --no-model on a short Claude Code session about a red build:

# Distilled by husk from ci-triage.jsonl (claude-code).
# 11 messages · confidence 0.33 · heuristic
# No durable user-supplied facts found worth carrying as knowledge.
# Heuristic distillation: no model was used. Review the persona before shipping.
# Review the persona before you serve this. Distillation is a draft, not an oracle.

apiVersion: husk/v1
name: triage
displayName: triage
description: "Act as the CI triager for this monorepo. CI on main is red again: find the faili"

model: sonnet

persona: |-
  Act as the CI triager for this monorepo.

  How you work:
  - Always read the failing job log before guessing, and never re-run a job without saying why.
  - Keep it short next time: the failing job, the first error, the likely commit, and the fix.

  You have a Linux computer. Run commands to check things rather than guessing,
  and show the command you ran when the result matters.

tools:
  - computer
  - files

# … then the two exchanges it kept as examples, and where it came from

The model is the one the chat used, the rules are the person's own words, and the tools are the ones it called. The computer's flavor, network policy and limits stay at their defaults until you set them.

husk distill prints its confidence and everything it could not work out, so a thin transcript announces itself instead of producing a plausible-looking persona nobody checks. Five worked examples live in examples/.

A bot runs on Husk's own engine, on any model, or on your own Claude Code or Codex, headless under your own login, with Husk's tools and nothing else:

engine:
  type: claude-code   # or codex; builtin is the default
  resume: <session>   # Claude Code: go on with that chat, on its computer

From a Claude Code chat, the console's Make this a bot sets this up when Claude Code is installed, so the bot is that chat going on; a Codex chat runs on Codex when Codex is installed. The husk.yaml reference has the details.

What is in it

• A machine, not an interpreter. Files, a shell, ports, a browser. /work survives the whole session, and persist keeps it past that.

• Five computer providers. docker, podman, ssh, fly, local. One interface; auto takes the highest available.

• Eleven model providers. Ollama, Anthropic, OpenAI, Google, Groq, DeepSeek, Cerebras, OpenRouter, Mistral, Together, LM Studio. Aliases resolve across all of them, so one husk.yaml runs on Opus or on a local Gemma.

• Four adapters. Discord, Slack, Telegram, webhook.

• No product telemetry. The installed runtime has no analytics, crash reporter, or version ping. The marketing and documentation websites use Vercel Web Analytics.

• No native modules. npm install finishes on Windows with no C++ toolchain. Node 20.10 or newer.

• Automated regression tests and archive smoke checks. npm test checks the packages; integration cases that need an unavailable provider are explicitly skipped.

How it works

Thirteen packages and five apps. Dependencies run downhill: core imports nothing from the workspace, every other package except console-assets (the built console, no code) imports it, and nothing imports @husk-ai/cli.

packages/
  core         contracts, husk.yaml schema, primitives
  runtime      computer providers: docker, podman, local, ssh, fly
  models       one surface over eleven model providers
  sessions     transcript importers + the distiller
  browser      Chromium lifecycle and CDP automation
  agent        tool-calling loop and built-in tools
  adapters     Discord, Slack, Telegram, webhook
  mcp          MCP server (stdio), and the shared Husk every chat forwards to
  server       control-plane API + bot host
  sdk          typed client for the control plane
  workspaces   named local workspaces: sources, results, export
  console-assets  the built console, served by server
  cli          the husk command
apps/
  console      dashboard: live terminal, files, browser
  desktop      the desktop app: corner, Computers window, notch, tray
  mac          Husk Bar, a macOS menu-bar companion
  docs         documentation site (Next.js + MDX)
  web          marketing site (Next.js + Three.js)

server declares @husk-ai/agent but reaches it through a dynamic import, so the control plane still starts on a tree where the agent was never built.

Four flavours choose the image: base, python, node, full. Husk publishes none of its own. base is debian:bookworm-slim, python is python:3.12-slim, node is node:22-slim, and full is Playwright's image, which already carries the twenty-odd shared libraries Chromium links against. An image you publish is an operating system you have promised to keep patched, and Debian and the Playwright team already do that better. Set HUSK_REGISTRY to point the lot at your own mirror.

Open interactively · SVG

Providers

auto walks the ladder and takes the highest rung that answers. Every rung is free except fly, which is metered, so auto leaves it out: pick it with --provider fly, or set HUSK_AUTO_FLY=1 to let auto choose it. An explicit --provider that turns out to be unavailable is an error, never a quiet downgrade to weaker isolation.

The local provider is not a sandbox. It stops accidents, not adversaries. husk doctor reports isolated: false for it, and the first MCP tool result says so again.

Security model

These hold in every mode, local included, and were re-checked against a live local and a live docker computer for 0.1.4: the command deny list, the environment scrub, output caps and process-tree kill. The computer tools also pass every result through redact() and every call through audited(). The workspace tools, hand-offs and job_wait are not audited.

The path jail is not one of them, and this paragraph used to say it was (#120). It is a control on local and ssh, where no container stands behind it, and there it confines the file tools rather than the shell. On docker, podman and fly the container is the boundary, so the file tools address the container's own filesystem: read_file /etc/passwd returns the container's copy, and /dev/shm takes a write. Neither reaches your machine, which is the point of running those providers — a jail inside a kernel boundary is largely redundant. Claiming one anyway is exactly what this section exists not to do.

What local does not give you is a kernel boundary. It shares your kernel, your network and your user account.

Husk's own tools and the browser's page loads refuse 169.254.169.254 even under network.mode: full, because full means the internet and not the cloud metadata service that hands IAM credentials to whatever asks. The RFC1918 ranges are the same problem one hop out, and so is loopback, except on Docker, Podman and Fly, where the computer has a loopback of its own. An operator who wants one of them names it in allow, where a reviewer reading the husk.yaml can see the decision.

That check is not a firewall. A shell command is not held to it: in a computer, curl http://169.254.169.254/ reaches the address wherever the computer can route to it. Only network.mode: none on Docker or Podman, which gives the container no network at all, stops a process. The security model has the matrix.

Address spellings are normalised first. 127.1, 2130706433 and 0x7f000001 are all 127.0.0.1 to curl, to Chromium and to Python's urllib, so a rule that only understands four dotted octets is not a rule.

Open interactively · SVG

Documentation

• Architecture — why it is shaped this way

• API reference — the control-plane HTTP contract

• Build contract — conventions every package obeys

• Security model — what is isolated and what is not

• Benchmark — Husk's browser tools against Playwright MCP, measured in CI

• Examples — five husk.yaml recipes and an MCP session demo

• Diagrams — 16 interactive architecture, dataflow and design diagrams

Status

Alpha, pre-1.0. The husk.yaml schema and the HTTP contract can still change between releases. What is known not to work today:

• On container providers the rendered browser needs the full flavor. CI drives the fifteen browser_* tools on Docker with full (Playwright's image, which carries Chromium's libraries). The slim flavors cannot run Chromium: their root filesystem is read-only, so the libraries cannot be added later. base (debian:bookworm-slim) has neither curl nor python3, so even browse fails there. Chats get full unless told otherwise (npx @husk-ai/mcp, husk mcp, remote access, and husk setup unless you answer no to the browser). husk up and a husk.yaml default to base: pass --flavor full, or set computer.flavor: full.

• fly is the least exercised provider. It has tests. It has far fewer real hours than docker and local.

Development

git clone https://github.com/Hotragn/husk.git && cd husk
npm install
npm run build
npm test

Contributing

Contributions are welcome. CONTRIBUTING.md covers the build contract, the test requirements, and how to add a provider or a model.

Contributors

Everyone who has shipped something here is on the contributors graph, which counts commits rather than asking anyone to remember to add a name.

License

Apache-2.0