First Run

Step 1: Install and open the app

GitHub Release: https://github.com/omnimind-ai/OpenOmniBot/releases

CNB, recommended for users in mainland China: https://cnb.cool/o.a/OpenOmniBot/-/releases

We provide two package variants. The in-app auto-updater will keep using the variant you selected. If you want to switch variants later, install the other package manually once to overwrite the current one.

For development builds:

./gradlew :app:installDevelopStandardDebug -Ptarget=lib/main_standard.dart # install the standard slim version
./gradlew :app:installDevelopOmniinferDebug -Ptarget=lib/main_omniinfer.dart # install the version with local model inference

After the first launch, the welcome page appears while the app initializes the Alpine system in the background:

Step 2: Complete the core permissions

Open the authorization page and prioritize these four permissions:

• background execution
• overlay permission
• installed apps access
• accessibility service

Then optionally add:

• notifications
• Shizuku
• storage and file-related permissions
• calendar and exact alarms

WARNING

If overlay and accessibility are missing, many failures look like “the agent understands the task but never actually acts.”

Step 3: Configure model providers

From settings, go to the provider configuration page and create a profile:

• choose OpenAI-compatible for OpenAI-style APIs
• choose Anthropic for Claude-style APIs

You will typically need:

• a profile name
• base URL
• API key
• available model IDs

Step 4: Bind scene models

Then open the scene model settings and bind appropriate models to the major scenes:

• Agent
• Operation
• Compactor
• Memory Embed
• Memory Rollup

If you only have one strong model at first, you can temporarily bind everything to it and refine later.

Step 5: Check the workspace environment

On startup the app prepares the workspace and default files automatically. You can also verify that the Alpine or runtime environment is initialized from settings.

Step 6: Test a minimal execution path

Use the chat page to validate one simple flow:

1. send a task that does not depend on a complex external service
2. watch for tool activity or execution cards
3. confirm that permissions, workspace access, or browser features are being called correctly

Good first test prompts

• “Summarize what actions are available on the current screen.”
• “Open the workspace and tell me what files are there.”
• “List the installed skills.”

Step 7: Enable advanced capabilities gradually

Once the base flow works, turn on more features one by one:

• local models
• skill store
• scheduled tasks
• MCP server
• remote MCP tools

That makes debugging much easier than enabling everything at once.

Next, read Model and Scene Setup.