> For the complete documentation index, see [llms.txt](https://docs.gotempest.app/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.gotempest.app/ai-and-automation/tempest-ai-assistant.md). # Use Storm AI in Tempest Storm AI integrates Apple Intelligence, Claude, GPT, Gemini, and self-hosted LLMs into your SSH terminal — explain commands, debug errors, generate one-liners. The **Storm AI** is an in-app chat panel that knows what's happening in your SSH session. Ask it to explain a stack trace, generate a one-liner, debug a failing systemd unit, or draft a migration script — and pipe the answer straight into the terminal. Think of it as Claude / GPT / Gemini sitting next to your shell, with full context. See [External Agents vs LLM Providers](/ai-and-automation/external-agents-vs-llm-providers.md) for a capability comparison. ## What it can see The AI sees only what you give it — by default that's: * The current command output you highlight * Files you explicitly attach (drag-and-drop into the chat) * Documents you pick from Tempest Drive * Your typed question It does **not** silently scrape your terminal history, your hosts list, or your keys. You stay in control of what reaches the model. ## Attach a file (Pro) Attach a document to a conversation and Tempest converts it to text on your own device before anything is sent. Supported formats: | | | | --------- | ----------------------------------------------- | | Documents | PDF, Word, PowerPoint | | Data | Spreadsheets, CSV | | Text | HTML, plain text in any encoding, including GBK | | Images | Sent directly to models that accept them | This works with every provider, not only with local agents. Long documents are measured against your model's actual context limit — counted correctly for Chinese — and the assistant tells you when a document was truncated, or when a scanned PDF contains no text to read. The assistant can also read the documents in your [Tempest Drive](/ai-and-automation/snippets-scheduled-runs.md) on request, attachments included. ## Connect your own MCP servers (Pro) Beyond the tools Tempest provides, the assistant can call tools from **MCP servers you run yourself** — a company knowledge base, an issue tracker, a database gateway. 1. Go to **Settings → AI → MCP Servers**. 2. **Import** and paste the `mcpServers` configuration you already use with Claude Desktop, Claude Code, Cursor, Windsurf or VS Code. Each row shows whether the server is currently reachable. Every call to one of these tools asks for your approval first, reads and writes alike — Tempest has no way to verify what a third-party tool actually does, so it never decides on your behalf. In Read Only mode they are withheld entirely. On iOS, servers launched as a local command cannot run and are listed as unsupported rather than hidden; the same configuration keeps working on your desktop. Not to be confused with the [Tempest MCP Server](/ai-and-automation/install-tempest-mcp-server-in-ai-clients.md), which is the opposite direction — Tempest offering *its* tools to an external AI client. ## What it can do * **Explain**: paste an error → get a plain-English diagnosis * **Generate**: "give me a one-liner to find files modified in the last hour" → get the command, optionally one-click "run in terminal" * **Investigate**: "the disk is full, show me the biggest directories" → get a `du` invocation, run it, AI sees the output and follows up * **Refactor**: paste a Bash script → get an idiomatic rewrite, side-by-side diff * **Bridge to coding agents**: combine with the **Use AI Agents** preset on a connection to drop straight into Claude Code, Gemini CLI, Codex CLI, or OpenCode on the remote host ## Choose your model Tempest doesn't lock you into one provider. Configure any of: * **Anthropic Claude** (Sonnet, Opus, Haiku) — recommended default * **OpenAI** GPT-4 / GPT-5 family * **Google Gemini** * **Apple Intelligence** — On device or Cloud (PCC) on supported Apple platforms * **Self-hosted** — point at any OpenAI-compatible endpoint (Ollama, vLLM, llama.cpp, your enterprise gateway) Configure a provider in **Settings → AI** and select its model in the assistant. External API providers need their API key; Apple Intelligence uses the native Apple framework and does not need a provider API key. Your AI settings — providers, API keys, external agents, MCP servers and speech endpoints — sync across your devices through your encrypted vault, so a provider added on the desktop is ready on your phone without retyping the key. ## Apple Intelligence Tempest can use Apple's Foundation Models through a native bridge in the **macOS desktop app** and **iOS app**. It is not available in Windows, Linux, Android or browser Web Mode. Syncing the provider setting to another device does not make that device eligible to run it. ### Enable it 1. On a supported Apple device, enable **Apple Intelligence** in macOS **System Settings** or iOS **Settings**, and allow the system model to finish downloading. 2. Open **Settings → AI → Providers** in Tempest (on iOS, open AI settings from **More**). 3. Choose **Add provider → Apple Intelligence**. 4. In the assistant's model selector, choose **Apple Intelligence → On device**, or **Cloud (PCC)** if available. 5. Check the provider availability status, then start a conversation. No API key is required. ### Choose the data path | Option | Where the model runs | Requirements and limits | | --------------- | ------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **On device** | Locally on the Apple device running Tempest | iOS/macOS 26 or later, eligible hardware/region, Apple Intelligence enabled and model ready. Tempest's configured context budget starts at 4,096 tokens. | | **Cloud (PCC)** | Apple's Private Cloud Compute | iOS/macOS 27 or later, a Tempest build with the required SDK support, eligible service/device and permission. The macOS app must have the PCC entitlement in its signed build. Tempest's configured context budget starts at 32,768 tokens. | The current provider defaults to **Cloud (PCC)** where the build offers it. If that option is unavailable, explicitly choose **On device**; adding Apple Intelligence alone is not a promise that PCC is usable. The native bridge checks availability for the selected model and reports the system's actual context limit where supported. **On device** keeps the model request local. **Cloud (PCC)** sends the request to Apple; it is not an offline mode. Both use Apple's native framework rather than a user-configured OpenAI-compatible HTTP endpoint. See [Apple Foundation Models](https://developer.apple.com/documentation/FoundationModels/) for Apple's framework and availability information. This local-processing statement applies to the model request, not every assistant action: SSH connections, remote tools, MCP servers and notifications can still use the network. Apple Intelligence models are text-only in Tempest; they do not directly accept image attachments as a vision model would. Responses arrive as completed answers rather than streaming token by token. ### Availability and context errors * **Turn on Apple Intelligence:** enable it in the system settings, then return to Tempest so availability is checked again. * **Device not eligible / unsupported OS:** use an eligible device and supported OS, or select another provider. * **Model downloading or unavailable:** allow the system setup/download to finish, then retry. * **Cloud requires iOS/macOS 27:** choose On device on a supported 26.x device. * **Missing PCC permission:** Cloud needs an appropriately signed/entitled app; use On device or a build that provides that permission. * **Context window exceeded:** start a new conversation, reduce attached content or send a smaller part of the terminal output. The system limit covers instructions, conversation and tools as well as the latest question. Apple Intelligence uses Tempest's normal tool approval and Read Only controls. Model availability does not bypass those controls or the entitlements required for particular Tempest tools. ## Privacy * Custom-provider requests go to the configured endpoint. Built-in hosted AI uses its service; Apple Intelligence follows the On device/PCC distinction above. In Web Mode, backend-mediated operations run on the deployment host, which must be trusted. * With a self-hosted model endpoint, model requests stay within the network hosting that endpoint; other tools or notification destinations can have their own data paths. * See [How Tempest Protects Your Privacy](/accounts-vaults-and-privacy/how-tempest-protect-your-privacy.md) for the full data-handling story. ## Configurable temperature Set the model temperature in Settings — lower for deterministic command generation, higher for creative refactoring suggestions. ## Coding agents preset The connection edit form has a **Use AI Agents** quick-config: pick **Claude Code**, **Gemini CLI**, **Codex CLI**, or **OpenCode**, give a working directory, and Tempest will `cd && ` on connect — perfect for jumping straight into a coding session on a remote dev box. See the dropdown next to the **Startup Command** field in any SSH host's edit form. ## See also * [Kubernetes Profiles](/terminal-and-operations/kubernetes-profiles.md) — pair the AI with `kubectl` for cluster troubleshooting * [Tempest Monitoring](/terminal-and-operations/tempest-monitoring.md) — when a gauge spikes, ask the AI what to do * [Tempest Drive — Saved Commands & Scheduled Runs](/ai-and-automation/snippets-scheduled-runs.md) — save AI-generated commands as reusable snippets * [Use Tempest in AI Agents (Desktop or CLI)](/ai-and-automation/install-tempest-mcp-server-in-ai-clients.md) — the other direction: your AI client calling Tempest's tools * [AI terminal](https://gotempest.app/ai-terminal) — the product page for this feature