Getting Started
Download and install Munk AI, configure AI, optionally enable Munk Test, then enter a project and run your first work item.
Follow this path from download to first real work. Munk AI currently supports macOS on Apple silicon (M-series) only.
1. Download and install
Get the desktop app from the website Install page: follow @iBoyCoder on X / Twitter for private preview access, or join the waitlist for email updates.
After install, open Munk AI and sign in.
2. Configure AI
Open Settings → AI and set your OpenAI-compatible provider:
- Base URL (for example OpenRouter)
- API key (your own model credentials; the key is never shown again after save)
- Default model (refresh the list, or type a custom model id)
- Click Save
Configure MCP, Skills, webhooks, and the rest later as needed. For account and repos, see Account and GitHub.
3. Optional: enable Munk Test (UI verification)
Do this only if you want AI to run device / UI verification (Deploy → Verify) inside the loop. You can code and run gates without it.
- Install and configure Test with Munk Test Getting Started (including the Test Agent API key)
- Back in Munk AI, open Settings → Test
- Confirm a resolved
munkbinary (usually leave blank for auto-resolve) - Click Connection test until serve is reachable and Verify is ready
- Then turn on Enable UI Verify
See Relationship with Munk Test for the product split.
4. Home: how you enter a project
After launch, Home offers four entry points:
| Entry | When to use it |
|---|---|
| New project | Start a local project from scratch (ready-to-use coding environment) |
| Open local | Point at an existing folder as a Codebase |
| Clone public | Clone a public repository locally |
| GitHub App | Clone or link a repo from your authorized GitHub App list |
The Recent list jumps back into projects you used lately.
For GitHub authorization and binding details, see Account and GitHub.
5. Workbench: the main surface at a glance
Once you open a project, you land in the Workbench — where you spend most of the day:
Read it by region:
| Area | Role |
|---|---|
| Left project list | Switch local Codebases |
| Work items column | Units of work for this project (requirements / bugs / Issues, and so on) |
| Center Session | Chat with the agent for the current stage; clarify goals, read reasoning, send messages |
| Top-right Workflow | Intake → Coding → Gates → Deploy → Verify → Delivery; failures can loop back to Coding |
| Bottom-right code / tree | Inspect related files; switch to Diff when needed |
Items that need a human decision show up in the top Inbox.
6. Minimum path: one work item end to end
- Create or open a project (any Home entry)
- Create a work item — think GitHub Issue, product requirement, or bug; or wait for GitHub / Feishu / DingTalk intake
- In Session, make the goal clear (the agent may ask follow-ups; answer and let it continue)
- Watch Workflow advance; on Gates / Verify failure it re-enters repair under Host rules
- When you must decide, open Inbox and act
You do not need MCP, Skills, webhooks, or UI Verify for the first pass. Success is: project → work item → chat → follow the flow.
7. Optional: follow progress away from the desk
Keep the desktop signed in and disable system sleep (screen off / lock are fine). Then use the website Remote UI to check progress and chat with agents. See Remote UI.