AI Settings and Write Controls
Who this is for
Organization Owners and Admins who configure AI Pilot's provider, safety controls, and granular write permissions.
What you will complete
Configure your AI provider and API key, set the master safety controls, and fine-tune exactly which write operations AI Pilot can propose.
Before you begin
- Owner role is required to change AI provider settings and API keys.
- Admin role is required to change agent controls and per-server access.
- You need an API key from your chosen AI provider.
Access the AI settings page
- Go to Settings in the left sidebar.
- Click AI Agent in the settings menu.
The AI Agent settings page has four sections: AI Provider, Agent Controls, Per-Server Access, and AI Memory.
Step 1: Configure your AI provider
- Under AI Provider, select your provider from the dropdown:
- Anthropic (Claude) — Claude Sonnet 4 (recommended), Claude Haiku 4 (fast), Claude Opus 4 (reasoning)
- OpenAI (GPT) — GPT-4o (recommended), GPT-4o Mini, o3, o4-mini, GPT-5.x series
- Google (Gemini) — Gemini 2.5 Flash (recommended), Gemini 2.5 Pro (reasoning), Gemini 2.0 Flash (fast)
- Ollama (Local) — Gemma 3, Llama 3.3, Mistral, DeepSeek R1, and others running locally
- Select a model from the Model dropdown. Models tagged "recommended" are the best balance of speed and capability. Models tagged "reasoning" are slower but better for complex multi-step plans.
- Set Max Sequential Steps Per Run — this controls how many tool steps the AI can take in one response cycle. The default is 100. Lower this if you want the AI to pause for review more often. Minimum is 1.
- Enter your API Key (or Server URL for Ollama). The key is masked and stored encrypted. Leave the field empty to keep a previously saved key.
- Anthropic keys start with
sk-ant- - OpenAI keys start with
sk- - Google keys are obtained from Google AI Studio
- Click Save AI Settings.
- Click Test Connection to verify the key works. A success or error message will appear.
Step 2: Configure agent controls
Under Agent Controls, you will see the following toggles. All are independent — disabling one does not affect the others.
AI Pilot (master switch)
When off, the agent cannot perform any actions on your infrastructure. You can still chat for questions and advice.
Read Access
Allows the agent to view server metrics, system info, alert history, and site listings. This is read-only — it cannot change anything.
File and Log Access
Allows the agent to browse site files and read server logs via SSH. Useful for debugging. It cannot modify or delete files with this permission alone.
Write Actions
Allows the agent to restart services and execute approved infrastructure plans. All write actions still require your explicit approval via an approval card.
Fine-tune Write Actions
When Write Actions is enabled, click Fine-tune Write Actions to control exactly which operation categories and sub-operations the AI can propose. Categories are:
| Category | Sub-operations |
|---|---|
| Server Operations | Provisioning, Power Control, Delete and Disconnect, Resize Disk, Package Management, Firewall, Services and SSH |
| Site Operations | Create and Clone, Delete Sites, File Editing, SSL Certificates, Cron and SFTP |
| Deploy Operations | Deploy and Rollback, Environment Variables, Repository Linking |
| Backup Operations | Create Backups, Restore Backups, Delete and Download, Offsite and Schedules |
| Database Operations | Create and Drop Databases, Database Users |
Toggle a category off to disable all its sub-operations. Toggle an individual sub-operation off to block only that one action while keeping the rest of the category enabled.
Auto Memory
Allows the agent to automatically save notes about your preferences and infrastructure setup as you chat. You can review and delete these memories at any time.
Opt out of AI Pilot learning corpus
By default, scrubbed lessons from your conversations (without your data, secrets, or IPs) are used to improve AI Pilot for all users. Enable this toggle to opt your organization out entirely.
Production Protection
When enabled, the agent cannot perform write actions on servers marked as production. Recommended for all organizations running production workloads.
Step 3: Save and verify
- Each toggle saves immediately when clicked. You will see a confirmation toast.
- To verify your settings, ask AI Pilot to perform a simple read action. Confirm it responds correctly.
- To test write controls, ask the AI to propose a low-risk action (such as checking a service status) and verify the approval card appears as expected.
What success looks like
- AI Provider shows "Configured" badge next to the API key field.
- Test Connection returns a success message.
- Agent Controls show the correct enabled/disabled state for each toggle.
- AI Pilot responds to chat messages and proposes actions only within the permitted operation scope.
Common errors and fixes
"Test Connection fails with 'Invalid API key'" Cause: The key is incorrect, expired, or belongs to a different provider than selected. Fix: Check the key in your provider's console. Ensure the provider dropdown matches the key type. Re-enter the key and save before testing.
"I changed the provider but the AI is still using the old one" Cause: The model field may still be set to a model from the old provider. Fix: After changing the provider, change the model to a valid option for the new provider and click Save AI Settings again.
"Fine-tune Write Actions is not visible" Cause: Write Actions toggle is off. Fix: Enable Write Actions first, then the Fine-tune section appears below it.
"I disabled a sub-operation but the AI still proposed it" Cause: Sub-operation settings take effect on the next request. An in-progress conversation may use cached settings. Fix: Start a new conversation in AI Pilot after changing sub-operation settings.
"Max Sequential Steps has no visible effect" Cause: The AI completed the task in fewer steps than the limit. The limit only prevents the AI from continuing past that step count. Fix: Lower the limit further if you want the AI to pause more frequently for multi-step operations.
Safety notes
- Disabling AI Pilot (master switch) immediately stops all pending AI operations in progress.
- The API key is encrypted at rest. Staff cannot view your key in plain text.
- Fine-tune Write Action settings apply to all members of your organization. Per-user controls are not available at this time.