5-minute quickstart
Install the desktop app, configure model credentials, create your first agent, and run a real conversation.
One path only: from nothing to a successful model call. No terminal required.
Install the desktop app
Open the download page and pick the package for your platform:
| Platform | File | Notes |
|---|---|---|
| macOS (Apple Silicon) | AgentKit-Studio-mac-arm64.dmg | Signed and notarized |
| Windows 11 (x64) | AgentKit-Studio-win-x64.exe | NSIS installer |
Apple Silicon only
There is no Intel (x64) build. Intel Mac users should use a managed cloud runtime or a supported environment.
On first launch, if macOS reports that the developer cannot be verified, open System Settings → Privacy & Security and choose Open Anyway. On Windows, use More info → Run anyway in the SmartScreen prompt.
Configure models and credentials
The credential guide opens automatically on first launch. If you dismissed it, go to Settings → Models and credentials.
- Select or add a model provider endpoint.
- Paste the API key.
- Pick a model.
Create your first agent
- Open the Agents panel.
- Choose More ways to create → Manual configuration.
- Complete the three-step form:
| Step | What you decide |
|---|---|
| Basics | Name, description, system prompt |
| Capabilities | Runtime type (defaults to DSH Native Harness), model, optional MCP / skills / knowledge base / memory |
| Policy and confirmation | Execution policy (defaults to workspace-write), approval rules, cloud memory toggle |
- Choose Create. This produces the first revision.
Send the first message
Select the agent in the list and type a request, for example "List the files in the current workspace and explain what they do".
You will see:
- Tool calls expand in order, each with its arguments and result.
- Token usage and duration for the turn at the bottom.
- If the agent changed files, a file change card appears on the right so you can diff each one.
Verify success
A successful run satisfies all three:
- A complete final answer is rendered (not an error summary).
- Tool calls report
completed; there is noRUNTIME_RUN_FAILED. - The conversation still exists after a refresh.
If any check fails, use Troubleshooting to look up the error code.