Spool

Spool

思簿 — a context hub for long-running projects.

Logo: a spool viewed from above, its thread pulling free — watch it assemble.

The app icon has an editable Icon Composer source; one command exports it for macOS, Windows and the website.

spoolapp.org · Download for macOS · Download for Windows (v0.6.1)

The site walks the whole loop in an interactive demo (English / 中文) — no install needed.

macOS: a signed, Apple-notarized .dmg . Windows: an x64 installer, not code-signed yet, still v0.6.1 until the next release.

Free, offline, no account.

At the moment you naturally produce a fragment of information — a good answer from an AI, a decision buried in an email, a link to a document, a half-formed thought — Spool lets you capture that fragment effortlessly, files fragments under a two-tier Workspace → Project structure, and packs any project into a paste-ready briefing on demand — so you can re-enter a project, or re-brief an AI, instantly.

Why it exists

LLMs do not remember your project. Every new conversation, you re-explain the context. Across multiple AIs, multiple tabs, multiple emails, spanning many days, a project's context gets shredded — and reassembling it falls on your memory.

Spool compresses "re-explaining" into "a single paste."

The core loop

   Capture  ──▶  Project ──▶  Pack  ──▶  (paste, re-enter)

▲ │ └───────────── days later ──────────────┘

• Capture — double-tap a modifier key (⌥ on macOS, Ctrl on Windows) and whatever is on your clipboard is written into the current project. Rides existing ⌘C / Ctrl+C muscle memory. No decisions at capture time. The confirmation overlay opens with the cursor already in a note box — type the thought that made you save this and press Enter, or click anywhere to skip. Your own note outlives the excerpt, and connected AIs treat it as the highest-signal line. The main window never has to come forward.
• Project — an append-only timeline of fragments. Two tiers only: Workspace (big topic) → Project (one piece of work). No infinite nesting. On open the feed lands at the newest blocks — they ARE "where you left off," no manual status note to maintain.
• Pack — one click assembles a paste-ready Markdown briefing of the project. Pure string assembly, no AI in the hot path, fully deterministic.

Status

v0.7.4 — released for macOS only. Windows stays on v0.6.1, the last release that shipped a Windows installer; its next installer will be built from the current release once it has been checked on a Windows machine. What changed: v0.7.4 · v0.7.3 · v0.7.2.

On Windows, capture is a double-tap of Ctrl instead of ⌥. What did not cross over is the macOS-specific work around focus handling and browser tab-title capture, so a Windows capture records the source application rather than the page. Linux is not implemented.

• Distribution — macOS: Developer ID–signed, Apple-notarized .dmg, straight from Releases. Not on the Mac App Store — sandboxing conflicts structurally with the global capture trigger.
• Distribution — Windows: v0.6.1, an x64 NSIS installer built by CI, not code-signed, so SmartScreen warns once before it will install.
• Website: spoolapp.org, with an interactive walkthrough of capture → pack → re-brief → MCP in English and Chinese.
• No auto-update channel yet: direct distribution means new versions are a manual download.

New in v0.5.0

• Windows. NSIS installer, tray icon that stays legible on a dark taskbar, closing the window goes to the tray and says so the first time, and the same MCP hookup as macOS. Capture is a double-tap of Ctrl, read through Raw Input rather than a keyboard hook — which is what lets it work while antivirus software is watching for hook-based keyloggers. Neither platform binds a key combination out of the box; add one in Settings → Shortcuts if you want it.
• Undo covers your own writes. A block you wrote can be undone, and in an empty composer ⌘Z / Ctrl+Z means "undo my last action" rather than nothing.
• Follow-up is a list of lines. Each line is either *standing* (watch this indefinitely) or *one-off* (close it once answered). An AI on MCP can propose a line and close one it answered; proposals wait on the review screen, and a standing line refuses to be closed — without that distinction one answer would silently switch off a long-running watch.
• Workspaces nest, projects sort alphabetically, the left rail supports multi-select, and a whole workspace can be packed into a real folder (INDEX.md plus one .md per project, sub-workspaces as sub-directories) that an AI opens a piece at a time. Nothing is summarized away to fit a budget.
• Moving to another machine. Settings → Advanced → *Moving to another machine* exports the whole library and imports it elsewhere. Import merges; it does not replace.
• Provenance on a block: source URL, retrieved-on date, review-by date, with reminders two months, one month and a week ahead. Corrections got a proper interface, and a corrected block keeps its original wording.
• Capture fixes: typing goes straight into the note box again, and the overlay is placed by the screen's usable area so the menu bar no longer eats its top edge.

New in v0.4.0

The optional external-AI routes below add no keys or accounts to Spool: they drive a CLI you installed and logged into yourself. (Spool's own API engine, added in v0.6.2, is the one route that does hold a key — see The API engine.)

• A CLI engine slot. If you already have Claude Code, the Codex CLI, or the Gemini CLI installed and logged in, Spool can run it as a local subprocess to follow up on things you asked it to watch, write a cross-project weekly review, or help draft the lines worth following. The network request happens inside that CLI, on your own account's quota — for this route Spool stores no API key, and its own process makes no HTTP request. Gemini does not run Follow Up.
• Visible runs and honest records. Follow Up runs in the right-hand rail; Weekly Review has its own screen, and completed runs for those two actions leave a record. A follow-up-goal draft stays only in its editor until you save or discard it. Intermediate progress appears when the CLI exposes it, and nothing that would change your existing notes is applied silently.
• Follow up. You write a few plain lines describing what to watch for; the engine searches the web against exactly those lines and files what it finds for your review. It stays quiet when there is no news — the one action that deliberately reaches the open web, and the only one whose web tools are switched on.
• Retirement and correction, instead of overwriting. You can mark a block as no longer valid, or point at the one sentence in an older block that a newer one corrects. Retired blocks leave the pack but stay in the library and stay searchable, and the pack says out loud that they were left out. An append-only log must never silently overwrite a fact.
• Annotations say who wrote them. A note an AI wrote through MCP renders as ai note: in a pack and can never be read as your own judgement — that distinction is the whole point of the authority header.
• A review queue for AI writes. An AI splitting one passage across several projects, or proposing a correction, lands in a queue you approve; it does not land in your timeline.
• Packs are context, full stop. The three "what should the AI do" task types are gone — a pack now carries your context and the reading instructions, and you state the task yourself in whatever chat you paste it into.

All twelve phases of the original implementation roadmap were landed in v0.3.0:

Design principles (non-negotiable)

1. Capture must be zero-friction — one keypress, no decisions.
2. Local-first, private by default — Spool's own process makes no network request, and its CSP forbids one structurally. Content reaches another program only through a hand-off you choose: paste a Pack, enable an MCP client, run a CLI action, or switch on the API engine. Every one of those is off until you turn it on; MCP's write side is a separate switch again. When the API engine is on, the request is made by a local subprocess Spool starts (spool-ai), with your own key and quota — the boundary is the process, not a promise.
3. A project is a log, not a chat — append-only, time-ordered, quiet.
4. Retrieval is deterministic — pack and search never call AI or the network.
5. AI is a librarian, not an author — anything an AI files through MCP is attributed, append-only, and can never overwrite what you wrote by hand.
6. Exactly two tiers of structure — no infinite nesting.

The full product constitution, rejected ideas, and the feature filter are in PLAN_EN.md §2.

Stack

• Tauri 2 desktop shell, with a second non-activating overlay window for capture confirmations
• React 18 + TypeScript (strict mode) on Vite (multi-page build)
• Tailwind CSS for layout; design tokens in CSS variables
• Zustand for state
• SQLite via tauri-plugin-sql, FTS5 with the trigram tokenizer
• MCP server (spool --mcp, stdio, default OFF): the AI surface — 28 tools, 15 read (list/search/dedup/pack, a compressed block's original text, a block's pictures, a linked folder's status) plus 13 consented, attributed write tools. A Follow Up or Weekly Review run that Spool starts itself is served 26: everything except proposing a project split and marking a linked folder as synced
• CLI engine slot: claude, codex, or gemini runs as a local subprocess for Follow Up, Weekly Review, follow-up-goal drafting, compression, and the out-of-date check — detected on disk, never bundled, never given a key

Building from source

Requirements: Node 20+, Rust toolchain (stable), Tauri 2 system dependencies (see the Tauri docs for your OS).

npm install

npm run tauri dev # dev wi