Set up AI chat

WebF1 Pad has a full chat mode beside your notes. AI chat is off by default; you bring your own provider and keys.

Enable chat and add a provider

1. Open Settings (the gear button, or View → Settings) → AI Chat.
2. Tick Enable AI chat. The Notes/Chat switch appears in the title bar.
3. Add a provider:
   • Ollama - preset included; base URL http://localhost:11434 for locally running models. Nothing leaves your machine.
   • OpenAI-compatible - any endpoint speaking /chat/completions (OpenAI, OpenRouter, LM Studio, vLLM, …).
   • Anthropic - direct Claude API access.
4. Paste the API key (or a file:/path/to/key reference) and add the model ids you want. Auto-discover queries the provider’s /models endpoint and merges the list for you.
5. Click Test, then Save.

Chatting

• Switch to Chat in the title bar. New chats start with the provider and model you picked.
• The session sidebar lists your chats newest-first (the most recently worked-on chat on top, so resuming an older chat moves it back up). The panel button in the chat header collapses or reopens the sidebar at any window width; on narrow screens it opens it as a drawer instead.
• Rename a chat from the sidebar’s pencil button (or double-click the title), or click the title in the chat header. A manual rename is never overwritten by the automatic title suggestions.
• Attach a note with the paperclip (or File → Transfer Note to Chat), or attach a file - text, code, PDF or an Office document - with the same button or by dragging it onto the composer.
• Paste images straight into the composer - vision-capable models receive them.
• The Context selector (All / 40 / 20 / 10) limits how much recent history is sent to the model - the full chat stays visible locally.
• Chats are restored by the app automatically. To keep or share one, the chat offers Save as note, Save to cloud, Export PDF and Print; saved chat notes reopen as chats and can be backed up like other notes.

Chat titles (optional)

After the first reply in a new chat, a small LLM call proposes a short descriptive title for the sidebar entry. Settings → AI Chat → Chat titles turns this off, or picks a dedicated Title model from any of your providers (a searchable picker, the pick remembers its provider). Leave it empty to reuse the web-search Router model (if set), else the chat’s own model. One tiny call per chat, never applied to chats you renamed yourself or to saved chats.

Saved prompts

Reusable prompts live in the chat sidebar’s Prompts list:

• Click a preset to run it as a new chat.
• Use the add button to create one (a name, the prompt text and, optionally, a provider and model), and the pencil to edit it.
• The Command Palette (Ctrl+Shift+P) also lists them as Run prompt: …, so you can start one without opening the sidebar.

Tool calling

For tool-capable models, tick Tools on the provider (or per-model). The chat can then:

• save_note - write its answer into a new note (say “summarize this into a note”, or use the Save as note button).
• read_active_note - read the note you currently have open for context.
• search the web - if enabled (below), decide mid-conversation when a search would help and read full pages.
• call your MCP servers - see Give the chat more tools with MCP servers. This includes WebF1 Pad’s own MCP server: while it is enabled in Settings → MCP Servers, the chat can list/read/update your open notes, read chat sessions and answer “how do I …?” questions from the app’s How-To documentation.

The Tools chip in the composer flips tool use on or off for the current chat.

Tool rounds planning

Settings → AI Chat → Proactive tool rounds planning tells every chat how many tool rounds it may use and asks the model to estimate its needs before it starts work. If the model expects to need more than the limit, it asks you for a raise up front; you only see the approval dialog when it asks, and you can decline.

Provider diagnostics

If a provider misbehaves (empty replies, auth errors), Settings → AI Chat → Diagnostics → Log AI provider diagnostics writes each chat round’s provider, model, HTTP status and a truncated payload sample to <data dir>/logs/ai-diagnostics.log. The log can contain your prompts and note contents (via tool calls), so review it before sharing it; API keys are never logged. Open logs folder and Clear log sit right under the toggle, which applies immediately.

Completion sound

A short chime plays when an AI reply finishes streaming - including when it ends in an error, so you always hear that something happened. Pressing Stop is silent. Turn it off in Settings → AI Chat.

Web search (optional)

Settings → AI Chat → Web search enables grounded answers:

• DuckDuckGo works with no configuration and no API key.
• SearXNG uses your own instance if you self-host one.
• Tavily and Exa are LLM-focused search APIs, they need a free API key (setup links and free-tier details in Web search backends).
• Tool-capable models can decide when a search would help while chat tools are enabled. Other models show a per-chat Web switch and use a small routing step to decide whether to search. Search failures never break the chat.
• Optionally pick a Router model in Settings → AI Chat → Web search: a searchable provider + model picker (the same mechanism as the new-chat defaults). The pick remembers its provider, so the router can use a small/fast model from a different provider than the chat itself.

Everything runs from the app - no third-party script or service in the middle.