Assistance is optional and provider-dependent
AI can help explain terminal output or prepare an action, while supported voice workflows can turn speech into input or read text aloud. These tools depend on the provider, device permissions and capabilities enabled in your release.
Autocomplete starts with local history and command completions. When a confident local completion is available, it does not need a model request. Enabled model suggestions use the configured AI connection, with request pacing and local fallback if the provider fails. Do not assume that an encrypted terminal connection also makes a prompt private from the AI provider receiving it.
Configure an AI connection
- Open Settings → AI and review the available connection modes.
- For My API key or a self-hosted endpoint, first enable I'm Advanced user in Settings → Help & Diagnostics if those controls are hidden.
- Choose My API key under Inference source, then select the intended Server preset, API provider, API base and Model. Presets include Ollama, LM Studio and Custom self-hosted server; start the selected server before connecting.
- Enter an API key when authentication requires one. Authentication → None is available only for an explicitly selected local or self-hosted server; it does not send saved provider or account keys.
- Review the provider's data handling and costs. Where offered, use Discover models to read the model catalog. Select a discovered model or enter one manually, then use Verify inference to send a small model request. Catalog discovery alone does not verify inference, and the verification request can incur provider charges.
- Use Save settings, then test with non-sensitive input and review the result before using it on a real server.
Saving writes the key through this workspace's credential store, outside ordinary settings JSON. Encrypted credential backups may include it; hosted-cloud workspaces save it in their encrypted cloud document. Browser access can require the provider's browser-origin support or an approved HTTPS Gateway. Do not work around a failed browser request by sending the key to an unknown proxy.
LeapTerm account is offered where account features are enabled and requires the configured account service. A provider's charges and limits are separate from LeapTerm account credits and task spending limits. Local completions do not verify provider access or remaining quota.
Installed coding-agent support
Installed host coding agent is a separate inference source. The current selector offers supported Claude Code and Gemini CLI inference; Codex and OpenCode are marked inference unavailable. The current Gemini inference path requires Gemini CLI 0.21.3 on the connected host. Installing either unavailable command-line tool does not enable its inference integration. You can still use an installed tool through an ordinary authorized terminal; that is a different workflow.
Review context and voice permissions
Depending on the selected feature, prompts, selected terminal context or speech may be processed by the configured service. Review that service before sharing passwords, private output or recordings. A terminal screenshot or pasted image can reveal more than the question you intended to ask.
Open the terminal's Voice settings control, or Voice & STT Settings… in its voice menu. The current selectable backend is Device / Browser Speech, when the device or browser reports support; other speech providers are marked unavailable. Choose an available voice or the system default and use Save Voice Settings. Unlock the workspace vault if settings cannot be read or saved.
Grant microphone access only for a voice workflow you intend to use. Device or browser speech can depend on an online service, so check its data handling before assuming audio stays on the device. Stop the active voice operation when finished; revoke operating-system or browser permission when it is no longer needed.
A suggestion is not a safety guarantee
Inspect generated commands, paths and destinations before execution. Models can misunderstand an environment or produce incorrect actions. For automated file and command access, read agents and tasks and connected assistants, where explicit permissions matter beyond ordinary text suggestions.