Open Computer Use

MCP server that gives any LLM its own computer — managed Docker workspaces with live browser, terminal, code execution, document skills, and autonomous sub-agents. Self-hosted, open-source, pluggable into any model.

> Online demo: **lab.widemoat.ai** — Open WebUI with Computer Use already set up, sign in with GitHub or Google. (More ways to try it below.) The old `chat.yambr.com` address redirects here and will keep doing so.

>

> Where this project is going. Open Computer Use set out to answer one question — can an LLM be given a real computer safely enough to be useful? It answered it, and it is used in production. That result led us somewhere else: Wide Moat, an enterprise AI platform that runs inside a company's own perimeter. It is a different product, not a rewrite of this one, and it is currently developed in private.

>

> Practically, for you:

>

> - This repository keeps working. It is maintained — fixes, dependency and security updates, and support for the Open WebUI versions it targets. It is not abandoned and not deprecated.

> - Its pace of new features slows down. Our attention has moved to the platform, and that is honest to say up front rather than to leave you guessing from commit dates.

> - The licence promise stands. FSL-1.1-Apache-2.0: use it, fork it, self-host it, redistribute it — and every release converts to Apache-2.0 two years after publication, whatever we do next. Nothing here can be taken back from you.

>

> Worth watching if you like this project: a sandbox integrated natively into Open WebUI, rather than bolted on through a filter and a tool, is one of the things being built on the platform. Try the hosted lab at lab.widemoat.ai or read more at widemoat.ai.

>

> If any of this looks useful, a ⭐ on the repo really helps — thanks!

What is this?

An MCP server that gives any LLM a fully-equipped Ubuntu sandbox with isolated Docker containers. Think of it as your AI's computer — it can do everything a developer can do:

• Execute code — bash, Python, Node.js, Java in isolated containers
• Create documents — Word, Excel, PowerPoint, PDF with professional styling via skills
• Browse the web — Playwright + live CDP browser streaming (you see what AI sees in real-time)
• Run Claude Code — autonomous sub-agent with interactive terminal, MCP servers auto-configured
• Use 13+ skills — battle-tested workflows for document creation, web testing, design, and more

Built for production multi-user deployments. Tested with 1,000+ MAU. Each chat session runs in its own isolated Docker container — the AI can install packages, create files, run servers, and nothing leaks between users. Works seamlessly across MCP clients: start with Open WebUI today, switch to Claude Desktop or n8n tomorrow — same backend, no migration.

Key differentiators

Works with any MCP-compatible client: Open WebUI, Claude Desktop, LiteLLM, n8n, or your own integration. See docs/COMPARISON.md for a detailed comparison with alternatives.

Live browser streaming

File preview with skills

Frontend design — landing page rendered live in the browser tab

Presentations — custom design system, not the default white template

Build your own skills — package recurring work into reusable functions

Data → chart with analysis

Claude Code — interactive terminal in the cloud

Sub-agent dashboard — monitor and control

See docs/FEATURES.md for architecture details and docs/SCREENSHOTS.md for all screenshots.

> Pro tip: Create skills with Claude Code in the terminal, then use them with any model in the chat. Skills are model-agnostic — write once, use everywhere.

> Multi-CLI sub-agent runtime (v0.9.2.1+): The sub-agent dispatch supports Claude Code (default), OpenAI Codex, and OpenCode (with OpenRouter / qwen / DeepSeek / 75+ providers). Flip `SUBAGENT_CLI=claude|codex|opencode` in `.env` — see docs/multi-cli.md for the worked OpenCode + qwen3-coder + OpenRouter recipe.

Architecture

> Looking ahead: a Kubernetes-friendly architecture with object-storage-backed user data and squashfs-packaged skills is being designed in docs/future-architecture/. Docker Compose remains the primary supported path.

Ways to try it

OAuth only — no email/password, no SMS. On `lab.widemoat.ai` models are bundled as a free convenience. The hosted MCP endpoint is offline during the transformation; see docs/CLOUD.md.

Quick Start

git clone https://github.com/Wide-Moat/open-computer-use.git
cd open-computer-use
cp .env.example .env
# Edit .env — set OPENAI_API_KEY (or any OpenAI-compatible provider)

# 1. Start Computer Use Server (builds workspace image on first run, ~15 min)
docker compose up --build

# 2. Start Open WebUI (in another terminal)
docker compose -f docker-compose.webui.yml up --build

Open http://localhost:3000 — Open WebUI with Computer Use ready to go.

> Note: Two separate docker-compose files: `docker-compose.yml` (Computer Use Server) and `docker-compose.webui.yml` (Open WebUI). They communicate via `localhost:8081`. This mirrors real deployments where the server and UI run on different hosts.

Model Settings (important!)

After adding a model in Open WebUI, go to Model Settings and set:

Without `Function Calling: Native`, the model won't invoke Computer Use tools.

What's Inside the Sandbox

Skills

13 built-in public skills + 14 examples:

14 example skills: web-artifacts-builder, copy-editing, social-content, canvas-design, algorithmic-art, theme-factory, mcp-builder, and more.

See docs/SKILLS.md for details.

MCP Integration

The server speaks standard MCP over Streamable HTTP. Point any MCP client at your own deployment.

• Self-hosted: `http://localhost:8081/mcp`. Quick sanity check:

curl -X POST http://localhost:8081/mcp \
    -H "Content-Type: application/json" \
    -H "X-Chat-Id: test" \
    -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

Full self-host integration guide (LiteLLM, Claude Desktop, custom clients): docs/MCP.md. The per-chat system prompt rides six redundant MCP-native channels (tool descriptions, `/home/assistant/README.md` in the sandbox, `InitializeResult.instructions`, `resources/list` for uploaded files, plus an HTTP `/system-prompt` endpoint for legacy integrations) — full map in docs/system-prompt.md.

Configuration

All settings via `.env`:

Custom Skills & Token Management (optional)

By default, all 13 built-in skills are available to everyone. For per-user skill access and custom skills, deploy the Settings Wrapper — see settings-wrapper/README.md.

Personal Access Tokens (PATs): The settings wrapper can also store encrypted per-user PATs for external services (GitLab, Confluence, Jira, etc.). The server fetches them by user email and injects into the sandbox — so each user's AI has access to their repos/docs without sharing credentials. The server-side code for token injection is implemented (`docker_manager.py`), but the Open WebUI tool doesn't pass the required headers yet. This is on the roadmap — if you need PAT management, open an issue.

MCP Client Integrations

The Computer Use Server speaks standard MCP over Streamable HTTP — any MCP-compatible client can connect. Open WebUI is the primary tested frontend, but not the only option.

Open WebUI Integration

> **Open WebUI** is an extensible, self-hosted AI interface. We use it as the primary frontend because it supports tool calling, function filters, and artifacts — everything needed for Computer Use.

Compatibility: This build is strictly built and verified against Open WebUI 0.11.0. The first 3 segments of our build version (`v0.11.0.X`) always match the Open WebUI base version it targets. If you run a different Open WebUI version, pick the Open Computer Use build whose first 3 version segments match yours — e.g., for Open WebUI 0.8.12 use a `v0.8.12.Y` build.

Why not a fork? We intentionally did not fork Open WebUI. Instead, everything is bolted on via the official plugin API (tools + functions) and build-time patches for missing features. This means you can use stock Open WebUI 0.11.0 with this build (the version that the first 3 segments of our build version `v0.11.0.X` match) — just install the tool and filter. Patches are applied at Docker build time; strongly recommended — 4 of them affect user-visible UX (artifacts panel, preview iframe, error banners, large tool-result handling). Pulling `ghcr.io/open-webui/open-webui` directly skips all of them — see Required setup when embedding Open WebUI for the full checklist.

Running Claude Code through a corporate gateway (LiteLLM, Azure, Bedrock)? See docs/claude-code-gateway.md for the three-path operator recipe.

The `openwebui/` directory contains:

• tools/ — MCP client tool (thin proxy to Computer Use Server). Required — this is the bridge between Open WebUI and the sandbox.
• functions/ — System prompt injector + file link rewriter + archive button. Required — without it the model doesn't know about skills and file URLs.
• patches/ — Build-time fixes for artifacts, error handling, file preview. Optional but recommended — improves UX significantly.
• init.sh — Auto-installs tool + filter on first startup. Optional — you can install manually via Workspace UI instead.
• Dockerfile — Builds a patched Open WebUI image with auto-init. Optional — use stock Open WebUI + manual setup if you prefer.

How auto-init works

On first `docker compose up`, the init script automatically:

1. Creates an admin user (`admin@open-computer-use.dev` / `admin`)

2. Installs the Computer Use tool via `POST /api/v1/tools/create`

3. Installs the Computer Use filter via `POST /api/v1/functions/create`

4. Configures tool and filter valves (`ORCHESTRATOR_URL=http://computer-use-server:8081` — internal URL for server↔server, seeded into both Valves)

5. Marks the tool public-read (access grants for both `group:*` and `user:*` wildcards) — so non-admin users see the tool in their workspace

6. Marks the filter both active and global (two separate toggles: `/toggle` and `/toggle/global`) — active-but-not-global is silently inert and a common manual-setup mistake

7. Merges `{function_calling: "native", stream_response: true}` into `DEFAULT_MODEL_PARAMS` via `POST /api/v1/configs/models` — every model gets the right defaults without per-model Advanced Params clicks

A marker file (`.computer-use-initialized`) prevents re-running on subsequent starts.

> Note: Open WebUI doesn't support pre-installed tools from the filesystem — they must be loaded via the REST API. The init script automates this so you don't have to do it manually.

Manual setup (if not using docker-compose)

If you run Open WebUI separately, you need to manually:

1. Go to Workspace > Tools → Create new tool → paste contents of `openwebui/tools/computer_use_tools.py`

2. Set Tool ID to `ai_computer_use` (required for filter to work)

3. Configure Valves: `ORCHESTRATOR_URL` = internal URL of your Computer Use Server (`http://computer-use-server:8081` for Docker compose)

4. Open the tool's ⋯ → Share menu and set access to Public (grants read to both `group:*` and `user:*` wildcards) — otherwise only your admin account sees the tool and non-admin users get an empty tool list with no error

5. Go to Workspace > Functions → Create new function → paste `openwebui/functions/computer_link_filter.py`

6. Enable the filter: toggle Active *and* toggle Global in the Functions list — these are two separate switches, and active-but-not-global means the filter loads but is never applied to chats

7. In your model settings, set Function Calling = `Native` and Stream Chat Response = `On`. Or set them globally once in Admin → Settings → Models → Advanced Params (`function_calling: native`, `stream_response: true`) — that becomes `DEFAULT_MODEL_PARAMS` for every model.

The docker-compose stack handles all of this automatically.

Required setup when embedding Open WebUI into your own stack

If you run Open WebUI outside the stock `docker-compose.webui.yml` — your own compose, Kubernetes, Portainer, or a downstream repo — there are four traps that will silently break Computer Use. All four hit us in production. Check in this order.

Step 1 — Build the image from `openwebui/Dockerfile`, don't pull upstream

Pulling `ghcr.io/open-webui/open-webui:vX.Y.Z` gives you a stock image without any of this repo's patches. Four of them are critical for UX:

Only `CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS` keeps working on an upstream image (it's a stock Open WebUI env) — which creates a false "everything is configured" feeling.

Use `build:` in your downstream compose, mirroring `docker-compose.webui.yml:11-15`:

services:
  open-webui:
    build:
      context: ./openwebui   # path into this repo
      dockerfile: Dockerfile
      args:
        OPENWEBUI_VERSION: "0.11.0"
    image: open-webui-with-cu-patches:latest   # local tag, do not pull

Verify the patches are baked into the running container:

docker exec open-webui bash -c \
  'grep -rl "FIX_ARTIFACTS_AUTO_SHOW" /app/build/_app/immutable/chunks/ >/dev/null \
   && echo "patches applied" || echo "MISSING — you are on upstream image"'

The `FIX_ARTIFACTS_AUTO_SHOW` JS comment marker is injected by `fix_artifacts_auto_show.py` at build time as a version-stable identifier — it does not depend on minified Svelte variable names, which change with every Open WebUI release.

Step 2 — No build-arg required for preview URL detection (host-agnostic since v0.9.2.0)

`fix_preview_url_detection` is now fully host-agnostic. The injected JS reads the origin directly from the matched URL at runtime (`_pm[1]` captures the full `https://host:port` prefix), so the patch requires no build-time host configuration. The `COMPUTER_USE_SERVER_URL` build-arg has been removed from `openwebui/Dockerfile`.

No action needed — the patch works automatically regardless of whether you use `localhost:8081`, a public domain, or Docker internal DNS. The preview iframe src is always reconstructed from the URL the model wrote into the message, which in turn comes from the server's `PUBLIC_BASE_URL` env var.

Verify the patch is applied:

docker exec open-webui bash -c \
  'grep -rl "FIX_PREVIEW_URL_DETECTION" /app/build/_app/immutable/chunks/ >/dev/null \
   && echo "patches applied" || echo "MISSING — fix_preview_url_detection not baked in"'
# → should print "patches applied"

Step 3 — Two URL settings, two roles (public vs internal)

v4.0.0: the old "three `FILE_SERVER_URL` places that must match" footgun is gone. There are now only two places and two distinct roles — public (browser-reachable) vs internal (Docker-local). The `COMPUTER_USE_SERVER_URL` build-arg was removed in v0.9.2.0 — `fix_preview_url_detection` is now host-agnostic (see Step 2).

⚠️ Do NOT point `ORCHESTRATOR_URL` at your public domain. It technically works, but every MCP request then goes browser→CDN→Traefik→container. Any hiccup in that chain kills the stream mid-tool-call and the user sees `MCP call failed: Session terminated`. Stay inside the Docker network.

The filter no longer has a public-URL Valve at all — it reads the public URL from the server's `X-Public-Base-URL` response header and caches it alongside the prompt. One public knob, one internal knob.

See also docs/openwebui-filter.md.

Step 4 — Four env vars on the `open-webui` container

Copy-paste into your downstream compose `environment:` block:

services:
  open-webui:
    environment:
      # --- Computer Use required env vars (read by build-time patches) ---
      - CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS=200
      - TOOL_RESULT_MAX_CHARS=50000
      - TOOL_RESULT_PREVIEW_CHARS=2000
      # Internal URL of the Computer Use server — seeded by init.sh into both
      # Tool and Filter Valves, and read by the fix_large_tool_results patch.
      # Same Docker network: use the service DNS name.
      - ORCHESTRATOR_URL=http://computer-use-server:8081

> Note: the last three are no-ops if the image is upstream ghcr.io — they need `fix_large_tool_results` from Step 1.

Step 5 — Filter must be global, tool must be public-read

Open WebUI has two separate switches for each function (`is_active` and `is_global`) and two required grants for each tool (`group:*` + `user:*`). The stock `init.sh` does this for you; manual / custom deployments commonly miss one side and then spend hours wondering why "everything is installed but nothing works."

Verify against the database (Postgres used by the stock stack; see `docker-compose.webui.yml:53`):

# Filter flags — expect (t, t):
docker exec  psql -U openwebui -d openwebui -c \
  "SELECT is_active, is_global FROM function WHERE id='computer_use_filter';"

# Tool grants — expect TWO rows (group|* and user|*, both 'read'):
docker exec  psql -U openwebui -d openwebui -c \
  "SELECT principal_type, principal_id, permission FROM access_grant WHERE resource_id='ai_computer_use';"

For SQLite-backed Open WebUI deployments, swap `psql` for `sqlite3 /app/backend/data/webui.db` with the same SQL.

Step 6 — Verify everything at once

# 1. Image has patches (marker-based — version-stable across Open WebUI releases):
docker exec open-webui bash -c \
  'grep -rl "FIX_ARTIFACTS_AUTO_SHOW" /app/build/_app/immutable/chunks/ >/dev/null \
   && echo OK || echo MISSING'

# 2. Preview URL detection is host-agnostic (no build-arg needed since v0.9.2.0):
docker exec open-webui bash -c \
  'grep -rl "FIX_PREVIEW_URL_DETECTION" /app/build/_app/immutable/chunks/ >/dev/null \
   && echo "patches applied" || echo "MISSING — fix_preview_url_detection not baked in"'
# → should print "patches applied"

# 3. Env vars reached the container:
docker exec open-webui env | grep -E 'CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS|TOOL_RESULT_|ORCHESTRATOR_URL'

# 4. Tool+Filter Valve (Session-terminated trap) — Admin UI is simplest:
#    Workspace → Tools → ai_computer_use → Valves → ORCHESTRATOR_URL
#    Admin → Functions → computer_link_filter → Valves → ORCHESTRATOR_URL
#    → both must be http://computer-use-server:8081 (internal URL, Docker service DNS),
#      NOT your public domain.

# 5. Server env (baked into system prompt AND returned to filter via header):
docker exec computer-use-server env | grep ^PUBLIC_BASE_URL=
# → must be a URL your browser can reach (e.g. http://localhost:8081 for local dev).

# 7. Filter is ACTIVE *and* GLOBAL (see Step 5):
docker exec  psql -U openwebui -d openwebui -c \
  "SELECT is_active, is_global FROM function WHERE id='computer_use_filter';"
# → expect (t, t). Two 't's, not one.

# 8. Tool is public-read with both wildcards (see Step 5):
docker exec  psql -U openwebui -d openwebui -c \
  "SELECT principal_type, principal_id, permission FROM access_grant WHERE resource_id='ai_computer_use';"
# → expect TWO rows: (group, *, read) and (user, *, read).

> After rebuilding the image, do a hard reload in the browser (Cmd+Shift+R / Ctrl+Shift+R). Otherwise it keeps the old cached JS chunks and you'll think the fix didn't work.

Symptom → which step is wrong

Security Notes

> Production tested with 1000+ users on Open WebUI in a self-hosted environment. For public-facing deployments, see the hardening roadmap below.

Current model

• Docker socket: The server needs Docker socket access to manage sandbox containers. This grants significant host access — run in a trusted environment only.
• MCP_API_KEY: Set a strong random key in production. Without it, anyone with network access to port 8081 can execute arbitrary commands in containers.
• Sandbox isolation: Each chat session runs in a separate container with resource limits (2GB RAM, 1 CPU). On Docker Compose, containers use the standard runtime (runc) and share the host kernel. The Kubernetes Helm chart defaults to rootless Podman — user namespaces plus AppArmor, and no privileged container — and can opt into Kata Containers for a hypervisor-grade boundary. On Compose, switch to gVisor (see roadmap). Containers have network access by default.
• POSTGRES_PASSWORD: Change the default password in `.env` for production.

Known limitations

• Unauthenticated file/preview endpoints: `/files/{chat_id}/`, `/api/outputs/{chat_id}`, `/browser/{chat_id}/`, `/terminal/{chat_id}/` — accessible to anyone who knows the chat ID. Chat IDs are UUIDs (hard to guess but not a real security boundary).
• No per-user auth on server: The MCP server trusts whoever sends a valid `MCP_API_KEY`. User identity (`X-User-Email`) is passed by the client but not verified server-side.
• Credentials in HTTP headers: API keys (GitLab, Anthropic, MCP tokens) are passed as HTTP headers from client to server. Safe within Docker network, but use HTTPS if exposing externally.
• Default admin credentials: `admin@open-computer-use.dev` / `admin` — change immediately in multi-user setups.

Security roadmap

We plan to address these in future releases:

• [ ] Per-session signed tokens for file/preview/terminal endpoints (replace chat ID as auth)
• [ ] Server-side user verification via Open WebUI JWT validation
• [ ] HTTPS support with automatic TLS certificates
• [ ] Audit logging for all tool calls and file access
• [ ] Network policies for sandbox containers (restrict egress by default)
• [ ] Secret management — move credentials from headers to encrypted server-side storage
• [ ] gVisor (runsc) runtime — optional container sandboxing for stronger isolation (like Claude.ai)

Ideas? Open a GitHub Issue. Want to contribute? See CONTRIBUTING.md or email developer@widemoat.ai.

Development

# Build workspace image locally
docker build --platform linux/amd64 -t open-computer-use:latest .

# Run tests
./tests/test-docker-image.sh open-computer-use:latest
./tests/test-no-corporate.sh
./tests/test-project-structure.sh

# Build and run full stack
docker compose up --build

Contributing

See CONTRIBUTING.md. PRs welcome!

Community

• Free online demo: lab.widemoat.ai — hosted by the maintainers
• Issues & Ideas: GitHub Issues
• Contact: developer@widemoat.ai

License

This project uses a multi-license model:

• Core (`computer-use-server/`, `openwebui/`, `settings-wrapper/`, Docker configs): Functional Source License, Version 1.1, Apache 2.0 Future License (FSL-1.1-Apache-2.0). Free to use, modify, fork, redistribute, and self-host internally. Each release automatically converts to Apache 2.0 two years after publication. Offering a hosted or embedded service that competes with our paid version(s) requires a commercial agreement.
• Our skills (`skills/public/describe-image`, `skills/public/sub-agent`): MIT
• Third-party skills: see individual LICENSE.txt files or original sources.

Attribution required: include "Open Computer Use" and a link to this repository.

See NOTICE for details. For third-party dependency licenses (PyMuPDF AGPL, Anthropic Skill License, Apache 2.0 bundles, etc.), see THIRD-PARTY-LICENSES.md.