Harness UI
A terminal and browser workbench for working on real projects with agents.
Harness UI is the Harness playground for individuals and trusted small teams. Use it to work on real projects while experimenting with models, instructions, tools, Skills, and execution environments. Ask an agent to explain a codebase, edit files, run checks, or delegate a focused investigation.
The terminal offers a personal coding-agent workflow. The browser adds shared conversations and drafts, live execution, files, Git changes, terminals, and configuration editing. Both use the same application and agent foundation; neither requires SDK code or a Service deployment.
Install and start
Install the a13n-harness-ui command with uv:
uv tool install a13n-harness-ui
cd your-repository
a13n-harness-uiFor the browser workbench, run:
a13n-harness-ui webuiOpen the login link printed by the server. The browser includes guided setup; see Use the browser for collaboration, authentication, native host access, and server lifecycle.
The terminal supports macOS, Linux, and Windows. The installed application does not require Node.js or a repository checkout. If the command is missing from PATH, run uv tool update-shell and open a new terminal.
On first launch, setup guides you through:
- Connect a model: use a supported subscription login or an API key. Existing compatible account stores can be reused.
- Select model settings: choose the offered model and, where applicable, service tier. You can tune reasoning and context later.
- Choose execution permissions: Full Control uses your host account; Sandbox requires working local isolation and does not silently fall back.
- Send your first prompt: setup saves editable files, then opens chat. No model request is made until you send a prompt.
Try a bounded first task:
Explain this repository's main entry point and tests. Do not modify any files.Full Control runs commands with your host account's filesystem and network access; it is not a sandbox. Sandbox requires supported Linux/macOS isolation; built-in Windows execution is Full Control only. See execution permissions.
Installation and upgrades covers source development and dependency updates. Setup covers login, cancellation, and advanced choices.
Where is my configuration?
Run this in your shell:
a13n-harness-ui config path
a13n-harness-ui config show --format json
a13n-harness-ui config validateOr type /config in chat. The default root file is ~/.a13n-harness-ui/a13n-harness-ui.yaml. Models and Agents live in sibling directories, not inside that file:
| Change | File or action |
|---|---|
| Default Agent, display, built-in tools | a13n-harness-ui.yaml — root reference |
| Model, endpoint, credentials, reasoning, context | models/*.yaml — Model reference |
| Instructions, tools, MCP selection, children | agents/*.yaml — Agent reference |
| Global coding guidance | AGENTS.md beside the root YAML |
| Project-specific guidance | AGENTS.md in the working directory |
| Additional workspace directories | projects/*.yaml — Project reference |
Start with common configuration recipes for copyable edits, or the configuration guide for all root fields and precedence. --config PATH selects another configuration tree; it does not merge it with the default tree.
Terminal controls
| Task | In chat |
|---|---|
| See available commands | /help |
| Switch Agent | /agent |
| Select and remember a Model per Project | /model |
| Change reasoning or priority service | /thinking, /fast |
| Change execution permissions | /environment |
| Inspect current configuration and usage | /status |
| Browse saved conversations | /resume |
| Add guidance while work is running | Type a message and press Enter |
| Cancel work | Ctrl+C or /cancel |
See Use the terminal for attachments, approvals, questions, history, and recovery. Add reusable resources outside chat with a13n-harness-ui add model or a13n-harness-ui add agent.
Go further
- Customize: Agents and subagents, Models and authentication.
- Connect tools: MCP and extensions, native tools and Web providers.
- Work across directories: Environments and Projects.
- Script a task or diagnose a failure: Automation and troubleshooting.
- Build another interface: Embed the Python App or use the HTTP API.
Sharing and execution boundaries
Share a WebUI instance only with trusted collaborators: they share credentials, configuration and accessible files, not separate participant permissions. Native Host Files and terminals are on by default; --no-share-computer turns them off independently of the Agent's execution mode.
Closing a browser does not stop an active Run; stopping the application does. Saved conversations can resume from the last checkpoint, but input or output since then may be lost. Use Service for managed identities and recoverable runs.