AI Command Bar

The AI command bar is available from a terminal or the AI sidebar. Describe the outcome you want in plain language, then review what Zync proposes before you run it.

Press Ctrl+I on Windows and Linux, or Cmd+I on macOS, to open the AI sidebar. The shortcut is configurable in Settings and Shortcuts.

Tip

Start with a precise outcome and a scope. For example, show the process using port 8080 is easier to review than fix my server.

Ask mode and Agent mode

The sidebar has two modes:

Mode	What it does	What you approve
Ask	Returns a command, explanation, or conversational answer.	You choose whether to copy or run the returned command.
Agent	Plans and performs a multi-step task with tools such as run_command, read_file, write_file, and list_files.	Review the plan and each safety checkpoint. Stop the run whenever needed.

Ask mode is a command helper. Agent mode is an autonomous workflow that can change files or execute commands on the selected local or SSH session. They share provider configuration, but Agent mode has a much larger action surface.

Caution

Treat Agent mode as an operator with access to the selected session. Check the target host, plan, file diffs, and command output before approving actions, especially on production systems.

First run

1. Open Settings > AI.
2. Turn on Enable AI Features if it is disabled.
3. Choose a provider and model. Ollama can run without an API key.
4. For a cloud provider, unlock the Local Vault when Zync asks and save your API key.
5. Open a terminal, press Ctrl+I or Cmd+I, and choose Ask or Agent.

The provider and model pills are also available in the sidebar. Ollama model choices come from the models installed at its URL. Cloud model choices are loaded from the provider when possible, with a fallback list for common models.

Providers

Zync supports these six providers:

Provider	Key	Default model	Where the request goes
Ollama	No	The first available local model	The configured Ollama URL, http://localhost:11434 by default
Google Gemini	Yes	gemini-2.0-flash	Google Gemini API
OpenAI	Yes	gpt-4o	OpenAI API
Claude	Yes	claude-sonnet-4-5-20251101	Anthropic API
Groq	Yes	llama-3.3-70b-versatile	Groq’s OpenAI-compatible API
Mistral	Yes	mistral-large-latest	Mistral’s OpenAI-compatible API

Provider defaults can change as providers retire models. Use the model picker to select a model currently returned by the provider.

Ollama

Install Ollama from ollama.com, start its local service, and pull at least one model:

ollama pull llama3.2

If Zync reports that Ollama is not running, start Ollama and check the URL in Settings > AI. If the selected model is missing, pull that exact model or select one shown in the model picker. A custom Ollama URL can point to a reachable Ollama service, but requests and any context sent to that service are no longer limited to this computer.

Cloud providers

Create a key with the provider, then paste it into the API key field in Settings > AI. Zync sends the key through the Rust backend to the provider request. Zync does not provide usage credits, and provider charges or limits are controlled by your provider account.

API key storage and migration

Cloud provider keys are stored as encrypted API-token records in the on-device Local Vault. They are not kept as plaintext in settings.json or exposed as a normal renderer setting. Unlocking the vault is required to save or resolve a key.

Older Zync versions could place keys in secrets.json or under ai.keys in settings.json. On migration, Zync copies a key into the Local Vault only when a vault record for that provider does not already exist, then removes the known legacy provider entries. An existing vault value wins over a legacy value.

Note

AI provider keys are local-only credentials. Zync excludes them from individual credential sync, bulk sync, and full Vault backup exports. Locking the vault prevents new key resolution. An AI request or Agent run that already resolved a key can finish.

What context is sent

For an active session, Zync can attach:

• the connection type, local or SSH
• the detected operating system and shell
• the current working directory
• the last 20 terminal output lines when terminal context is enabled
• explicitly attached terminal or file context

Terminal output is redacted by default before it is added to an AI request. Redaction reduces accidental disclosure, but it is not a guarantee that a provider can infer no sensitive information. Review the selected provider and the attached context before sending a request.

With Ollama at a local URL, the request stays on the local Ollama path. With Gemini, OpenAI, Claude, Groq, or Mistral, the prompt and attached context go to that provider. Zync does not route these requests through a Zync-hosted AI service. Agent tool output, file content, and diffs can also become part of the provider conversation while an Agent run is active.

See the security notes for the broader vault and provider trust boundary.

Using Ask mode

1. Select Ask in the sidebar.
2. Type a request such as find the process using port 8080.
3. Press Enter to stream the response.
4. Read the answer, command, explanation, and SAFE, MODERATE, or DANGEROUS classification.
5. Copy the command or choose Run in terminal when you are satisfied with it.

Ask mode keeps query history per session. Use the up and down arrows in the input to reuse recent queries. Clear the visible conversation from the sidebar when you want to remove that session’s AI history.

Using Agent mode

1. Select Agent in the sidebar.
2. Describe one bounded task, such as inspect why the API service is failing and show me the smallest safe fix.
3. Review the generated plan before approving execution.
4. Watch tool calls, output, file diffs, and checkpoints as the run proceeds.
5. Answer ask_user checkpoints explicitly. Use Stop to cancel the run.

Agent runs are scoped to the selected connection, or to the local machine when no SSH host is selected. Completed runs are retained as Brain sessions on the device so you can review the walkthrough and action log.

Safety classifications

Ask results use three risk labels:

• SAFE: normally read-only inspection such as ls, cat, ps, or df.
• MODERATE: changes that may be reversible, such as creating a directory or making a Git commit.
• DANGEROUS: destructive or high-impact operations such as recursive deletion, filesystem formatting, or destructive database statements.

The label is guidance, not a substitute for reading the command. Agent mode adds explicit checkpoints around actions that need confirmation.

Keyboard reference

Key	Action
Ctrl+I or Cmd+I	Open or close the AI sidebar
Enter	Submit the current request
Escape	Close the sidebar when focus is in the AI surface
Up and Down	Move through query history

Change the AI shortcut in Settings > Shortcuts > AI Command Bar.