Your First Agent Task
Use the AI assistant to find, install, and launch an app — entirely through natural language.
1. Open the Agent
Navigate to the Dashboard and click the sparkle icon, or select Agent from the sidebar. You'll see the agent hero screen:
What do you want to do today?
What would you like to do? Start typing or choose an example
Quick Actions
Choose an LLM model from the dropdown. Your last-used model is remembered automatically via CRDT sync.
No download needed for cloud API backends. For fully offline use: any Ollama model works — qwen3:4b (2.4 GB) is a good default.
Click Install app to pre-fill the prompt. Great for your first time.
Example Questions
Not sure what to ask? Try one of these:
- "Can you install an app for me?"
- "What are the most popular apps people run now?"
- "Help me set up image generation"
- "What models do I have downloaded?"
- "Debug why my app crashed"
- "Generate a manifest for my Python project"
2. Your First Prompt
Type your first request:
Find and install a local chatbot app
The status bar at the bottom turns pink and shows Thinking... with a clock icon. The agent is reasoning about your request.
What You Can Ask
| What You Type | What the Agent Does | Tools Called |
|---|---|---|
| "Install a chatbot" | Searches catalog, picks best match, installs | list_apps → install_app |
| "Find the best image gen for 8 GB" | Checks GPU, searches apps, recommends | gpu_diagnose → list_apps → hardware_status |
| "Why is ComfyUI crashing?" | Reads crash logs, checks GPU, diagnoses | app_diagnose → crash_log_debug → gpu_diagnose |
| "Edit my workflow to use a different model" | Reads file, patches checkpoint reference | file_read → file_edit |
| "What models do I have?" | Scans all model locations | model_list → hardware_status |
| "Package my Python app" | Analyzes repo, generates manifest | manifest_generate |
3. Watch It Work
As the agent works, you'll see several types of UI elements:
I'll search for chatbot apps available in the catalog...
- Search app catalog for chatbot apps
- Compare options and recommend best match
{"query": "chatbot", "tag": "llm"}
Found 3 apps: open-webui (4.5/5), sillytavern (4.3/5), text-generation-webui (4.1/5)
Thinking…
Open WebUI is the highest rated and most popular...
Status Bar
The status bar at the bottom cycles through states as the agent works:
- Thinking — LLM is generating a response
- Calling list_apps... — a tool is executing
- Ready — waiting for your next message
25 of 32 tools are safe and auto-run (all read-only tools). You only see approval prompts for the 7 tools that modify your system.
How Tool Execution Works
Each tool call follows a 6-step flow:
- Your message enters the conversation
- The LLM reasons about what to do (Thought:)
- It selects a tool and provides arguments (Action: + Action Input:)
- Permission check — safe tools auto-run, others need your approval
- The tool executes with a timeout
- The result feeds back into the LLM as an Observation
This cycle repeats until the agent reaches a final answer or hits the 25-iteration limit.
4. Approve Actions
When the agent calls install_app, you'll see a permission modal with the tool name, full arguments, and a risk description. You choose Approve or Deny.
Safety Levels
25 tools — auto-run, no approval needed
list_apps, app_status, app_logs, gpu_diagnose, gpu_test_pytorch, model_list, model_info, hardware_status, file_read, list_files, search_files, search_code, workflow_analyze, manifest_validate, manifest_generate, manifest_fix, install_debug, model_manage, crash_log_debug, app_manifest_debug, app_diagnose, model_manifest_validate, model_manifest_generate, model_cache_diagnose, web_search
7 tools — requires your approval before running
install_app, launch_app, stop_app, file_write, file_edit, run_command, model_download
run_command with destructive flags
rm, kill, chmod — extra caution advised
What to Check Before Approving
| Tool | What to Verify |
|---|---|
| file_write | File path — is it writing where you expect? |
| file_edit | old_string and new_string — is the change correct? |
| run_command | Full command text — watch for rm, kill, curl to unknown URLs |
| install_app | source_url — is it a trusted repository? |
| launch_app | app_id, port, and env variables |
| model_download | repo_id — is it the model you want? Check disk space. |
If You Deny
If you deny a tool call, the agent sees "User denied execution" as a tool result. It will continue reasoning and may try an alternative approach, or ask for your guidance. You can always say "try a different approach" or explain why you denied.
Always read run_command arguments before approving. The agent has shell access within the project directory.
5. Iterate & Explore
Example: Launch and Configure
Launch it
[launch_app {app_id: "open-webui"}]
Health check passed. Open WebUI is running at http://localhost:3000
What model is it using?
[app_status {app_id: "open-webui"}]
Connected to Ollama, using llama3.1:8b (4.7 GB)
Switch to something smaller
[hardware_status] → [model_list {filter: "gguf"}]
7.8 GB VRAM available. qwen3:4b (2.4 GB) is already downloaded — it fits well. I'll update the configuration.
Example: Debugging
ComfyUI won't start
[app_diagnose {app_id: "comfyui"}]
Process exited with code 1. Last crash: 2 min ago.
[app_logs {app_id: "comfyui", level: "error"}]
ModuleNotFoundError: No module named 'torch'
[install_debug {app_id: "comfyui"}]
Virtual environment exists but torch not installed.
[gpu_diagnose] → [run_command: pip install torch ...]
PyTorch was missing. I've reinstalled it with the correct GPU backend (CUDA 12.1).
Sessions persist via CRDT sync. Close the app, come back tomorrow — your conversation is preserved.
What's Safe to Explore
The agent can't break anything with read-only tools. Try these freely:
- "What GPU do I have?" — runs gpu_diagnose
- "List all my apps" — runs list_apps
- "Show ComfyUI logs" — runs app_logs
- "What models are downloaded?" — runs model_list
- "Check if my manifest is valid" — runs manifest_validate
- "How much disk space are models using?" — runs model_cache_diagnose
Anti-Patterns
- Don't ask the agent to read a 10,000-line file — use specific line ranges or ask for specific sections
- Don't approve
run_commandwithout reading the full command - Don't chain 10 requests in one message — the agent works better step by step
Key Takeaways
- Local by default — conversations stay on your device with Ollama or llama.cpp
- 32 tools total: 25 safe (auto-run) + 7 require confirmation
- Safe tools run automatically — no approval needed for reading or searching
- Always review tool arguments before approving writes and installs
- Sessions persist in CRDT — close and resume anytime
- For complex tasks, discuss before executing: ask "what would you do?" first