A computer for your AI agents. Run commands, search and manage files, execute code, and work with browsers and desktops through a self-hosted REST API. Connect it to Open WebUI, or use it with your own tools.
Documentation · Changelog · Report an issue
docker run -d --name open-terminal --restart unless-stopped \
-p 8000:8000 \
-v open-terminal:/home/user \
-e OPEN_TERMINAL_API_KEY=your-secret-key \
ghcr.io/open-webui/open-terminal:latestThe API is available at http://localhost:8000, with interactive docs at /docs. The volume keeps your files between container updates. Replace your-secret-key with your own key.
| Image tag | Includes |
|---|---|
latest |
Full toolkit: Python, Node.js, build tools, browser, desktop, and document tools |
agents |
Everything in latest, plus the latest Codex and Claude Code CLIs |
slim |
Minimal Debian environment with git, curl, and jq |
alpine |
Minimal Alpine environment with git, curl, and jq |
openshift |
Minimal image for OpenShift restricted non-root policies |
Use :agents to run codex or claude from a terminal, signing in with your own account. See the deployment guide for agent setup, image capabilities, custom packages, and updates.
Requires Python 3.11+.
pip install open-terminal
open-terminal run --host 0.0.0.0 --port 8000Or run without installing: uvx open-terminal run --host 0.0.0.0 --port 8000.
If no API key is configured, one is generated and printed at startup. For Docker, find it with docker logs open-terminal.
Caution
The API key grants command and file access. On bare metal, commands run with your user's permissions. Give access only to people and agents you trust.
- Open Settings → Admin → Integrations → Open Terminal.
- Add a URL reachable from the Open WebUI server and your API key.
- Enable the connection and select the terminal in chat.
Use the dedicated Open Terminal section to get the file browser and terminal panel. Compare connection modes, or see the connection guide for Docker networking and personal connections.
Install and sign in to the agent CLIs as the user running Open Terminal. In a compatible Open WebUI checkout, selecting this terminal adds its discovered models to the existing model selector as <agent>/<model>. Agent conversations run in the terminal's current directory, with streaming output, native tool/extension approvals, attachments, Stop, and conversation continuation.
| Agent | Native interface | Executable override |
|---|---|---|
| Codex | codex app-server --listen stdio:// |
OPEN_TERMINAL_CODEX_COMMAND |
| Claude Code | claude --input-format stream-json --output-format stream-json |
OPEN_TERMINAL_CLAUDE_COMMAND |
| Cursor | agent acp |
OPEN_TERMINAL_CURSOR_COMMAND |
| Grok | grok --no-auto-update agent stdio |
OPEN_TERMINAL_GROK_COMMAND |
| OpenCode | opencode acp |
OPEN_TERMINAL_OPENCODE_COMMAND |
| Cline | cline --acp |
OPEN_TERMINAL_CLINE_COMMAND |
| Gemini CLI | gemini --acp |
OPEN_TERMINAL_GEMINI_COMMAND |
| Pi | pi --mode rpc |
OPEN_TERMINAL_PI_COMMAND |
These adapters target current CLI protocols. ACP model discovery prefers session config options and also reads the native models field still advertised by current agents such as Gemini. Pi requires the current RPC agent_settled completion event and --session-id support. There are no hardcoded model lists. Log in through the CLI before using WebUI; the integration does not run interactive login flows. ACP agents execute their own filesystem and terminal tools in this environment. Pi follows its native tool policy; WebUI forwards extension confirmation dialogs but does not add a separate per-tool permission layer.
ACP history replay is suppressed when resuming an already saved WebUI conversation. If an agent does not advertise native resume, the next turn receives the saved conversation text instead. Agent-specific editor extensions beyond ACP's standard permissions and form elicitation are not implemented.
The same flow works through Tether's saved connection and outbound relay; the hosted server does not need to reach the local machine's IP. All three applications need the agent-capable changes. Provider CLIs and their login credentials stay in the terminal environment; no provider SDK is required in WebUI. Executable overrides above accept a path to the installed CLI.
The authenticated /agents catalog advertises installed providers and models. POST /agents/{provider}/sessions starts an owned process; /agents/sessions/{id} inspects or deletes it. Its /connect WebSocket carries native JSONL inside sequenced input/output frames with acknowledgements and replay. Sessions require X-User-Id and X-Session-Id ownership headers (single-user clients may omit X-User-Id). WebSocket clients send an initial { "type": "auth", "token": "<API key>", "cursor": 0 } frame. Agent turns use a single backend task and resume the provider's saved conversation on the next turn; branching/regeneration requires a new chat.
Configure Open Terminal with CLI flags, environment variables, or a TOML file. CLI flags take precedence.
| Guide | Covers |
|---|---|
| Configuration | Config files, browser sessions, storage, and file previews |
| Deployment | Docker variants, coding agents, custom packages, updates, and multi-user setup |
| Temporary public access | Cloudflare tunnels and troubleshooting |
| API reference | Interactive endpoints on your running instance |
Built-in multi-user mode separates workspaces inside one shared container; it does not isolate users and is intended for small, trusted deployments.
For separate containers per user, use Terminals. Production use requires an Open WebUI Enterprise License. See the security policy for the trust model.
Open an issue for bugs or setup problems. Report vulnerabilities privately through GitHub Security Advisories.
MIT. Bundled third-party tools retain their own licenses and terms.