drewnekota/cetus
One macOS app for Claude Code, Codex, and every agent runtime you use — scheduled runs, global hotkey launcher, per-run git worktrees, one review board.
安装Install
dsh plugin --profile web add github:drewnekota/cetus把 drewnekota/cetus 加入你的 DSH 配置（web profile）即可启用。
READMEREADME
Cetus
Turn your favorite agent runtime into an always-on desktop assistant.
Keep Codex, Claude Code, DeepSeek Harness, or the built-in runtime at the core. Cetus adds the desktop layer around it: summon it over any app, schedule work for later, and give it context from what has been on your screen.
Quick Launcher · Automations · Global Quick Reply · Screen Context
English · 简体中文
Join the Cetus community
Scan with WeChat to join the group chat and share feedback.
Four things Cetus adds to your agent
Quick Launcher — your agent, one hotkey away
Hold both ⌘ keys to summon your agent over any app. Cetus brings along the current screenshot, frontmost app, browser URL, and selected text as removable context chips, so you can ask in place instead of stopping to explain what you are looking at.
Automations — let it work while you are away
Turn any prompt into a one-off or recurring job (at / every / cron / daily). Each run keeps its chosen runtime and model settings, works in a fresh background conversation, and leaves the result ready for you to review.
Global Quick Reply — draft the answer without leaving the thread
Double-tap right ⌥ on any conversation — a team channel, an email, a support console — and Cetus reads what is on screen, then streams a draft into an editable panel. It follows the language and register of the thread and answers what was actually asked, so a message with three open questions does not come back as "sounds good". Press ⏎ to drop it into the input you were already in, or ⇥ to redraft the same screen on another runtime.
Screen Context — let it remember what you were doing
With screen context on, Cetus periodically captures frames, dedupes them, and OCRs them on-device with Apple Vision. Your agent can later recall what was on screen or search the history by text and app. Images and text stay on your Mac; capture is opt-in, with retention controls and an excluded-apps list.
Get started
The prebuilt app supports Apple Silicon and macOS 13 or later.
- Download the latest release.
- Open the DMG and move Cetus to Applications.
- Use the built-in Cetus runtime, or select an already installed and signed-in
claude,codex, ordsh(DeepSeek Harness) CLI. - Choose a workspace and give the agent its first task.
Claude Code, Codex, and DeepSeek Harness reuse their existing CLI login — there is no second account to configure. Building from source is documented under Development.
Early release: Cetus is under active development. Please open an issue if something breaks or if a workflow is missing.
Everything else
Keep the runtime you already trust
Use the built-in pi runtime, Claude Code, Codex, or DeepSeek Harness with the models, tools, and login you already have. Cetus translates each runtime into the same desktop workflow while preserving conversation context and background terminals across replies.
Pick a workspace, choose a runtime, optionally attach files or a screenshot, and send. For parallel coding tasks, enable per-conversation git worktrees so each agent edits an isolated checkout.
Review background work in one place
Every conversation is a card tracked across In progress · Needs review · Done. Automations, long-running tasks, and parallel solutions all surface here instead of getting buried in terminal sessions.
More ways to extend your runtime
- Persistent memory you and the agent both edit, injected into future turns
- Parallel solutions: fan one prompt into N candidate runs, then keep one and archive the rest
- Per-conversation git worktrees for isolated coding sessions
- Visual quick reply for drafting replies from the current screen without starting a full agent run
- Cetus Remote, an optional Tailscale-backed mobile companion for following runs and handling approvals
- Voice dictation and meeting memory, processed on-device
- Computer and browser control, with confirmation before consequential actions
- 30+ model providers, including Anthropic, OpenAI, Google, Bedrock, Ollama, LM Studio, and OpenRouter
Dictate from any app
Hold a hotkey from any app and talk — Cetus pops a floating equalizer HUD, transcribes on-device with Seed-ASR, and drops the cleaned-up text wherever your cursor is. The same stack as the in-app mic, but it follows you across the desktop.
This photo was taken with a phone because the floating overlay does not show up in screenshots.
Turn meetings into searchable context
Turn on meeting memory and Cetus quietly transcribes your calls into searchable notes — on-device, text only, no audio stored.
- Auto-detect — when another app grabs your mic (Zoom, Teams, FaceTime, Feishu…), Cetus starts a session and stops when the call ends. Nothing to press.
- Manual — global hotkey (default ⌘⇧M) for in-person meetings that auto-detect can't pick up.
- Both sides — your mic is you; system audio is everyone else, captured separately so the transcript knows who said what (macOS 14.2+; falls back to mic-only below that).
Transcription is 100% on-device via Apple's Speech framework — streaming, punctuated, segmented on natural pauses. While a session is live, a small floating pill (red dot + elapsed timer + stop button) sits at the top of your screen without stealing focus.
Those notes become context the agent can reach: ask "what did we decide about the launch date?" and Cetus searches your meeting history (search_meeting_history) — all local, nothing uploaded. Off by default; the master switch means Cetus never listens until you opt in. macOS only for now.
Stay in control
Each capability is opt-in. Computer & Browser control lets the agent drive your browser and Mac apps through numbered element lists (not raw pixels), with a confirmation step before anything consequential (sending, deleting, purchasing, submitting, authenticating) and a Stop button always in reach.
Why Cetus
General-purpose agents such as Codex, Claude Code, and DeepSeek Harness are already powerful. Cetus does not try to replace them. It gives them the parts of a desktop assistant that do not belong inside a terminal: an always-available entry point, ambient context from your Mac, continuity across sessions, background scheduling, and a place to review completed work.
The runtime remains the intelligence and execution engine. Cetus is the assistant layer around it. Choose the runtime for each task, add only the context you want, and make work that spans hours or days visible and inspectable.
That makes workflows practical that do not fit neatly into a terminal tab:
- Run an agent while you are away and review the result later.
- Compare independent solutions without colliding git changes.
- Carry project decisions and preferences into the next conversation.
- Connect coding work to the meetings, screens, and apps surrounding it.
Memory and Dreaming
The three factors describe a single moment. What makes an agent useful over time is whether it accumulates anything.
- Memory is context the agent writes back to itself — so the next session picks up where the last one left off instead of starting from scratch.
- Dreaming runs while you're idle: Cetus reflects on recent conversations and consolidates them into durable notes, turning raw history into preferences that persist. On by default.
Development
Requirements
- Node ≥ 20, pnpm, bun (for building the pi sidecar binary)
- Rust stable (
rustc,cargo) - Tauri prerequisites: https://v2.tauri.app/start/prerequisites/
- A
DEEPSEEK_API_KEY(or your provider of choice; pi auto-picks upANTHROPIC_API_KEY,OPENAI_API_KEY, etc.) - Optional: the Claude Code (
claude), Codex (codex), and/or DeepSeek Harness (dsh) CLIs installed and logged in, if you want them as conversation runtimes — Cetus reuses their existing login, no extra setup
First-time setup
pnpm install
# Build pi as a single-file binary into src-tauri/binaries/pi-<target>.
# Takes ~30s. Run once per dev machine; binaries are gitignored.
./scripts/build-pi-sidecar.sh
Run in dev
export DEEPSEEK_API_KEY=sk-...
pnpm tauri dev
Tauri launches the Next.js dev server (port 3000) and a window pointing at it. The pi sidecar is spawned automatically from the bundled binary.
Dev backdoor: PI_BIN
If you're iterating on pi itself, point at any pi build to bypass the sidecar:
export PI_BIN=/absolute/path/to/your/pi
pnpm tauri dev
This skips tauri-plugin-shell entirely and uses raw tokio::process::Command.
Build
./scripts/build-pi-sidecar.sh # if you haven't already
pnpm tauri build
Outputs .app / .dmg on macOS. A real multi-size icon set is required for tauri build (lives under src-tauri/icons/, regenerate with pnpm tauri icon <https://raw.githubusercontent.com/drewnekota/cetus/HEAD/path-to-1024px.png>).
Architecture
┌──────────────────────────────── Tauri window ──────────────────────────────────┐
│ │
│ Next.js (static export) Rust (Tokio + tauri-plugin-shell) │
│ ┌─────────────────────────┐ ┌──────────────────────────────────────┐ │
│ │ React UI │ invoke │ Tauri commands │ │
│ │ - ConversationList │ ───────► │ (list, new, switch, send, │ │
│ │ - Chat (text/thinking/ │ │ archive, set_model, │ │
│ │ tool cards), Composer │ ◄─────── │ extension_ui_respond, …) │ │
│ │ - ModelPicker (DeepSeek)│ event │ │ │
│ │ - DialogHost (ext UI) │ │ PiRpc: sidecar(plugin-shell) OR │ │
│ │ - chatReducer (deltas → │ │ PI_BIN(tokio::process) │ │
│ │ RenderedMessage[]) │ │ Store: SQLite metadata │ │
│ └─────────────────────────┘ └─────────────────┬────────────────────┘ │
│ │ stdin/stdout │
│ ▼ (LF-framed JSON) │
│ ┌──────────────────────────────────────┐ │
│ │ pi --mode rpc subprocess │ │
│ │ (bundled binary, any-model engine) │ │
│ └──────────────────────────────────────┘ │
└────────────────────────────────────────────────────────────────────────────────┘
- Conversations are pi
.jsonlsession files under<app-data>/sessions/. We index them (id, title, session_file, model, timestamps, archived_at) in SQLite at<app-data>/cetus.db. - Switching:
switch_session+get_messagesreplays history. One pi process for the app lifetime. - Streaming: pi emits
agent_start,message_updatewithassistantMessageEventdeltas, andtool_execution_*events. The frontendchatReducerfolds these into stableRenderedMessage[]indexed bycontentIndex, with atoolCallId → blockside-table to route execution updates. - Framing: strict-LF JSONL.
tauri-plugin-shelldelivers stdout in arbitrary byte chunks, so the reader maintains its own accumulator and emits one line per\n, stripping optional\r. Generic line readers that split on Unicode separators (Nodereadline) are non-compliant. - Sidecar packaging:
src-tauri/binaries/pi-<target>ships inside.app/Contents/Resources/.PI_BINenv var is the dev backdoor for iterating on pi. - CLI runtimes: conversations on Claude Code / Codex / DeepSeek Harness bypass the pi RPC entirely —
cetus-bridge::cli_agentkeeps a conversation-scoped Claude (or DeepSeek Harness) streaming session or Codex app-server thread alive, and a unit-testedEventTranslatormaps their events into the same PiEvent streamchatReduceralready consumes. Context and background terminals persist across turns via the vendor session/thread; optional per-conversation git worktrees isolate edits. - Extension UI: when a pi extension calls
ctx.ui.select()etc., pi sendsextension_ui_requestover the event stream. The frontendDialogHostrenders a dialog and replies via theextension_ui_respondTauri command. - Bridge: Cetus also intercepts known extension host tunnels and routes them to native handlers. See docs/bridge.md for the protocol, security boundary, and open-source extraction plan.
Reusable bridge packages
The host/extension bridge is factored into two standalone, provider-neutral packages you can depend on without pulling in the rest of the app:
cetus-bridge(Rust crate) — the product-light host runtime: JSONL subprocess RPC aroundpi --mode rpc, deterministic extension loading, host-tunnel classification, and injectableEventSink/TaskSpawnertraits. Tauri, app storage, and model-provider choices stay out of the crate — they live in app-side adapters (tauri_bridge.rs,app_event.rs,model_bridge.rs).examples/minimal_host.rsshows the smallest integration.@cetus/bridge-protocol(TypeScript) — the extension-side protocol: the sharedHOST_TUNNELSsentinels,callHost(),toolResult(), and host-tunnel types.
Both are MIT-licensed and carry no Cetus- or DeepSeek-specific code, so other agent hosts can reuse the same bridge. See docs/bridge.md for the protocol and security boundary.
Friendly links
- LINUX DO — A community of idealists.
License
MIT (matches pi).
相关推荐Related
-
nexu-io/open-design
🎨 Best DeepSeek Harness Design Plugin. The open-source Claude Design alternative. 🖥️ Local-first desktop app. 🖼️ Your…
-
esengine/DeepSeek-Reasonix
DeepSeek-native AI coding agent for your terminal. Engineered around prefix-cache stability — leave it running.
-
Molunerfinn/PicGo
:rocket: The Ultimate Image Uploader for Efficient Creators. Supports Obsidian, Typora, VS Code etc. and 60+ image hosti…
-
titanwings/colleague-skill
将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? It's giving rebirth era. Welcome to Digital …
-
anywhere-labs/deepseek-harness-desktop
为 DeepSeek Harness (DSH) 插件生态打造的现代化桌面端解决方案。万物皆「插件」，桌面本身也是「插件」。
-
tt-a1i/archify
Agent skill for beautiful, verifiable architecture, workflow, sequence, data-flow, and lifecycle diagrams—self-contained…
-
freestylefly/awesome-gpt-image-2
Prompt as Code | GPT-Image2 工业级提示词引擎与模板库，470+ 个案例逆向工程，20+ 套工业级模板，并提炼出Skills，持续更新中
-
YaoApp/yao
✨ All your agents and workspaces in one place, on every device you own. Track tasks on a board, accessible from desktop,…