Summon

Before you start: Summon is free — it just needs at least one configured AI model — run a local model (Ollama, LM Studio) with no key at all, or bring your own API key for a cloud provider (Anthropic, OpenAI, Google, AWS Bedrock, or an OpenAI-compatible cloud like Groq, OpenRouter, Mistral, DeepSeek, xAI or Together).

Opening Summon

• Click the glowing button in the bottom-right corner — it breathes quietly when idle and animates while Summon is working, with a badge counting in-flight tasks.
• Or type /summon in any note — your current block and selection travel along as context.
• Summon opens as a compact dock on the right edge; the expand button promotes it to a full-screen view. Close with Esc — the conversation survives, even as you switch notebooks.

Where Summon's answers go

A reply lives in the conversation by default. Two things move it into your notes:

• It writes as it works. When you ask Summon to build or edit something, it uses its notebook tools during the turn — you watch real blocks (headings, lists, checklists, tables) stream live into the editor, not a pasted wall of text. It can create a new notebook or write into the one you have open, and the newly-added blocks briefly flash so you can see exactly what changed.
• Pin a reply. Any message has a Pin action that drops its content into a notebook — new or existing — as a Section, for keeping an answer you didn't explicitly ask it to write.

Quick answers vs. Deep Research

By default Summon answers concisely — it does just enough tool work to reply, which keeps most look-ups fast (“what’s overdue?”, “when did I last mention @Sara?”). You don’t set this; it’s how Summon works out of the box.

When you want more, click the “Deep research a topic” quick-action chip. That arms Deep Research for your next question: a full multi-pass vault sweep before Summon writes, reading broadly and synthesising across many notes. Reach for it on prompts like “write a brief on everything I know about X” or “audit all my notes on the rebrand”.

Parallel research. For a broad ask, Summon can split the topic into several specific angles and research them simultaneously instead of one search at a time — you’ll see a status line like “Researching N angles in parallel…” while it works — then write the notebook from the combined findings. This is what makes Deep Research fast on genuinely broad questions.

What Summon can do

Summon isn’t a generic chatbot with your note pasted in — it has tools to work with your vault directly:

When the panel is empty, four quick-action chips greet you in the hero: Brief of the day, Prep for tomorrow, Deep research a topic, Create a new note. After each reply, follow-up chips suggest logical next steps — click one to keep going without retyping.

Composer Smart Tags

Type @ or # directly inside the Summon input to attach a Smart Tag to your prompt. The pill appears inline — and Summon automatically receives the entity’s full profile (attributes, linked notes) or the top matching vault passages as injected context. You can coin a brand-new tag on the spot if it doesn’t exist yet. This is the fastest way to scope a question: “Summarise everything I know about @Priya and #launch” — no need to mention them by name in plain text.

Asking about your vault

• Vault Q&A — ask “what do I know about X?” and Summon runs keyword full-text searches over your notes from multiple angles, then synthesises the answer right in the conversation — without creating a notebook you didn’t ask for.
• Scope cards — for open-ended research questions, Summon first shows a short scope card so you can pick the angle (or multi-select a few). It then researches exactly what you meant.
• Grounding badge — answers drawn from your own notes carry a small From your vault badge (or a passage count), so you can tell vault-sourced answers apart from the model’s general knowledge.
• RSS Memories are searchable too — articles you’ve pinned as RSS Memories (title, source and full text) are indexed when you save them, so Summon can pull from them as well.
• Your saved references, too — every reference in your library is full-text indexed, so Summon can pull from a paper you saved but never quoted. A fair share of its search results is reserved for references, so they’re never crowded out by your note matches.

Research output: citations & collections

Deep Research doesn’t just gather and summarise — it can author a properly-sourced notebook:

• Citations as it writes — when a claim comes from a source in your References library, Summon inserts a real citation block at that point (not a bare link), so the quote and its attribution live right where it’s used.
• A bibliography at the end — a research notebook Summon authors can close with a formatted References/bibliography section built from every source it cited, in your chosen citation style.
• Auto-filed into a collection — ask Summon to organize a notebook’s sources and it gathers everything it cited into a named Reference collection (reusing a similarly-named collection instead of creating a duplicate).

Writing & auditing in place

• Conversational edits in place — ask Summon to “expand on the decision loop” or “go deeper on the intro” and it inserts the new content after the relevant section, not at the bottom of the note.
• Gap audit — ask “what’s missing here?” or “is this comprehensive?” and Summon audits the open notebook and lists specific gaps as one-click chips you can fill immediately.
• Surgical task edits — ask Summon to add a task, tick a specific one off, or set a task’s due date and it changes just that item: every other row, its checked state and the list title stay exactly as they were (a due-date edit drops a tracked date pill and the note updates live). A built-in guard refuses any edit that would silently drop tasks, table rows or kanban cards.

Memory, steps & nudges

• Persistent memory (Brain button) — a memory panel behind the Brain button in the chat input shows everything Summon has learned about you across sessions. Add facts manually, or delete any it got wrong.
• Live agent steps — when Summon uses its tools, the steps stream live as collapsible “thinking” steps so you can follow its reasoning.
• Proactive nudges — Summon can surface cold notes you might want to revisit.

Assign work to Summon (background tasks)

Not every ask needs you to sit and watch. When a job can run on its own — draft a brief, chase something down, write a section — hand it off and keep working. Summon picks it up in the background instead of taking a live chat turn.

Three ways to assign, whichever feels natural in the moment:

• Type it inline — /assign draft a project brief for the rebrand. Everything after the command becomes the task.
• Select text, then /assign — highlight something you already wrote and delegate exactly that. Summon also gets the surrounding line for context.
• Just the block — run /assign on a line with no argument or selection, and the whole block becomes the instruction.

Summon drops an agent-task pill right where your cursor was — an inline chip that shows the task’s live status (queued → running → done, or failed) without you leaving the note. While the task is in flight, that block is locked so you can’t accidentally edit the instruction out from under the agent; deleting the pill cancels the task.

When a task assigned from a block finishes successfully, Summon stamps that source block #complete — so the line you delegated reflects its done state, and a native notification lets you know it’s ready.

The Summon Tasks view

Everything you’ve handed off lives in one place. Open it from the Summon Tasks row in the sidebar — it carries a small count badge of in-flight tasks and its background pulses gently while any task is running, so you always know something’s cooking without opening it.

The view leads with a one-line summary (e.g. “2 active · 1 needs approval · 12 done”), then an Active section and a windowed Recent list. Each task shows one of these statuses:

Done and cancelled tasks tidy themselves up automatically after 30 days, so the list stays manageable over time.

Gated actions pause for your approval. A background task can’t pop a dialog in your face, so when one wants to do something sensitive — write a file, run a shell command, or use a Mac bridge like sending mail or creating a calendar event — it stops and posts a Needs Approval card at the top of the Tasks view: “A background task wants to: …” with the concrete detail and inline Approve / Reject buttons. Nothing sensitive happens until you say so, and every decision is recorded. You choose which categories are even eligible — and how freely they run — in Settings → Summon → Capabilities (below).

Tip: assigning is the natural next step after Summon writes something for you. Turn a checklist of action items into real work — /assign each line and let them run while you move on.

Capabilities & permissions

Everything Summon can do beyond your notes lives in Settings → Summon → Capabilities — a two-pane page. Pick a capability in the left rail (grouped Access, Actions and Advanced) and configure it on the right. The guiding idea: reads are safe to leave on, and actions default to safe — nothing runs on your machine without the permission you grant.

Access — what Summon can read

Each is a plain on/off switch:

Actions — what Summon can do

Every action carries a single Never / Ask first / Allow control that is both its on/off and its permission level. “Never” blocks it, “Ask first” pauses for your OK each time, “Allow” runs it silently. Two of them add a Draft only setting. Turning an action on auto-enables the read domain it needs (switch that read domain off and its actions snap back to Never).

Terminal has a floor you can’t override: a built-in safety denylist — rm, curl, sudo, ssh and other destructive or exfiltration commands — is always blocked, even if you try to add one to the allowlist.

Advanced

Background tasks sets how many queued assigned tasks Summon runs in parallel. Higher clears a backlog faster but uses more of your AI provider’s rate limit; 2 is recommended.

Privacy reminder: when Summon answers a question about your notes, the relevant note content is sent to the AI provider you configured — that’s how the model reads it. Mac system tools (Calendar, Contacts, Mail) only send the data relevant to your specific request. Summon treats web pages, files and app content as untrusted and never follows instructions hidden in them. Nothing is ever sent anywhere when you’re not actively using Summon.