AI chat & inline AI
Press ⌘L to open the chat sidebar. Ask a question and the answer is grounded in the open canvas — retrieval runs over your notes, and citations appear as clickable chips that jump to the source note. Structured result cards show the retrieved notes above the prose answer. History is kept per canvas; Esc stops a streaming answer. (On Windows and Linux, read ⌘ as Ctrl.)
Where the model runs — and who decides
Local is the default, not a lock-in. Chat has four interchangeable engines; you pick per your needs, and the trust badge in the top bar always shows which one is active:
| Engine | What it is |
|---|---|
| Local (default, desktop) | A bundled model on your GPU or CPU (~3–5 GB one-time download). Zero network, zero per-token cost. |
| Apple Intelligence (iPhone/iPad) | Apple’s foundation models, system-managed, under Apple’s on-device + Private Cloud Compute privacy guarantees.* |
| Your own Ollama server | Point the app at an Ollama instance you already run — your models, your machine. |
| Cloud, your key (everywhere) | Any OpenAI-compatible endpoint + your own API key + model id — Claude- or GPT-class frontier models when you want more capability than local hardware gives. Always opt-in, behind a consent screen that names the endpoint host. |
* Requires an Apple Intelligence-capable device (A17 Pro-class iPhone or a supported iPad). On older devices the app explains why and walks you to the cloud-key option instead. You can’t load custom local models on iOS/iPadOS.
Three guarantees hold no matter which engine you choose:
- Swapping the engine swaps chat only — semantic search, embeddings, image captions, and voice transcription keep running on your device.
- Voice can never be routed to a cloud, even with cloud chat configured.
- Notes in the Confidential group are never sent to any cloud endpoint, regardless of provider settings.
On Android there is no local chat model — chat is your own cloud key or off (why). Local answers on desktop show a tokens-per-second badge, because speed depends on your hardware — that’s the honest cost of local AI.
Toggles in the chat header
- 🧠 Reasoning — per-canvas toggle that lets the model think longer before answering.
- 🌐 Web search — off by default, opt-in per canvas, and requires your own Brave Search API key. When on, queries go to that one allowlisted host and every request is counted in the in-app Network Audit.
Inline AI in the editor
Six editor actions run on the same local model: Continue, Rewrite, Summarize, Extract action items, Make callout, and Improve. Results stream in as ghost text — Tab accepts, Esc cancels and nothing is kept. Rewrite offers style presets before it runs.
Agentic edits — always behind a diff
Two larger operations exist, and both end at a review screen, never at silent changes:
- Ask AI to revise… (command palette) — describe the change, review the plan, then approve or reject a block-level diff.
- Consolidate notes (right-click a note) — merge overlapping notes into one, again behind a diff you approve.
On the roadmap: your agents and your notes
Two integrations are on the build roadmap — committed direction, not in the current build (the changelog will say so when they ship):
- MCP server — your agents, your notes. Smritsavant will expose your library to AI tools you already run — Claude, Codex-class coding agents, any MCP-capable client — as a local, permission-gated connection: read-only by default, write access behind a separate grant, Confidential notes never exposed, every access audited.
- MCP client — tools inside chat. Smritsavant’s own chat will be able to call tools you configure, under the same consent-and-audit rules as web search.
Setting up an alternate engine
Both live under Settings → AI & Search → AI Assistant → Advanced:
- BYO Ollama — enter the server URL + model name.
- Cloud, your key — enter any OpenAI-compatible base URL, your API key, and a model id. That covers the major providers and gateways directly, so Claude- and GPT-class models drop in with the key you already have. Enabling it shows a consent screen naming the exact endpoint host and what leaves your device; nothing cloud-bound happens before you accept, the trust badge flips to show the cloud host, and every request is counted in the Network Audit.