DeskWand Agent

Use any model from any provider.
Open source and local-first — your data never leaves your machine.

Website · Docs · Releases · 中文

---

Voice mode

Talk to it. Cut in whenever you want. It answers out loud, and every piece of speech work happens on your own machine.

Speak, don't type. Hold Right Option (macOS) or Right Alt (Windows) and talk. The transcript lands in the input box, cleaned up into written language.

Three quality tiers. Fast, balanced, or best-quality. The last one runs a 1.5 GB local model and keeps working offline — macOS Apple silicon today (about 3.8 GB RAM while speaking, released after 10 idle minutes).

Nine voices. Including Sichuan and Beijing dialects, with speed and speaking style you can change mid-conversation.

Interruption is a first-class feature. It hears the sentence you are finishing, not a timer — so a pause mid-thought doesn't cut you off, and a one-word filler ("嗯", "对", "好") doesn't throw away the answer.

---

Hand it the task. Watch it finish.

Four real jobs, every step visible.

Meeting notes → Word

Turns raw notes into decisions, owners and deadlines in a Word document you can send.

MEETING MINUTES · PROJECT NOVA
Weekly meeting · July 18 · 6 attendees · 45 minutes

── written in 1m 12s ───────────────────────────────────────────────

decisions
  · expand the canary to 10%        →  Wang Lin
  · backfill the Portal dashboard   →  Li Zhe
  · send the risk list to support   →  Chen Fan

Web research → sourced brief

Finds the sources, cross-checks them, and cites every claim it makes.

5 sources cross-checked
  reuters.com         ✓
  nvidia.com          ✓
  theinformation.com  ✓
  trendforce.com      ✓
  +1 excluded — data out of date

── finding ─────────────────────────────────────────────────────────

Data-centre GPU lead times have fallen back to 6–8 weeks, approaching
2023 levels [1][2], but HBM is still the bottleneck — next generation
volumes depend on SK Hynix's expansion [3][4].

Spreadsheet → the problem in one look

Opens the real workbook, works out the movement and the cause, writes the page.

Region         Q2      Q3     QoQ   Q3 revenue
East       12.84M  17.71M  +37.9%   ████████████
South       9.62M  10.48M   +8.9%   ███████
North      11.05M   9.38M  −15.1%   ██████
Southwest   4.18M   5.22M  +24.9%   ███

North fell because two contracts lapsed — 78% of the drop.

A bug report → tests that actually pass

Reads the repo, fixes the defect, runs the suite, and shows you the change.

  src/labels.ts
  function label(name: string) {
-   return name;
+   const text = name.trim();
+   return text || 'Untitled';
  }

  ✓ labels.spec.ts    ✓ form.spec.ts    ✓ api.spec.ts
  12 passed · 0 failed

+ branch fix/empty-label — waiting for your review

---

How it works

  your goal ──▶ plan ──▶ subagents + tools ──▶ result ──▶ a skill
                                                                │
 next task starts here ◀────────────────────────────────────────┘

---

Everything in the box

Agent core

Capability	What it does
Goal-driven execution	Give it a goal — it plans, uses tools, and keeps going until the job is done.
Subagents	Complex work splits into specialised agents that run in parallel.
Self-improving skills	A successful task can become a reusable skill, so the next one starts further along.
Web search built in	No separate key — search inherits the provider key you already set up. Dedicated backends (Brave, Exa, Tavily, Parallel, Perplexity) if you prefer.
Pi extensions & marketplace	Run Pi CLI extensions, install plugins, skills and prompt templates from the built-in marketplace.

Desktop & documents

Capability	What it does
Desktop native	Works with your real files, projects and environment — not a sandboxed upload box.
Agent-native Office	Ask for a report, a spreadsheet or a deck; the agent writes the actual Word, Excel or PowerPoint file.
On-device OCR	Reads text out of screenshots and scanned pages without uploading them.
Messaging channels	Reach your agent from Telegram, Discord, QQ, Slack, WeChat or Feishu.

Models, cost & privacy

Capability	What it does
Any model, any provider	Bring your own key — Anthropic, OpenAI, Gemini, DeepSeek, GLM, Kimi, MiniMax, OpenRouter, Ollama, or any compatible endpoint. Switch per session.
Standalone vision model	Pair a vision model with a text model, so non-multimodal models can still read images.
Usage and estimated cost	Priced from a public price table — by model, by day, by currency, for this machine only.
Local-first by default	Prompts, files and sessions stay on your machine; cloud sync is opt-in, not required.

---

Open source, local-first

DeskWand is MIT-licensed and built on the Pi Agent SDK. Bring your own model key: your prompts, files and sessions stay on your machine, and model calls go only to the provider you chose. Two things do leave the machine — an anonymous usage ping and the update check — and both can be switched off in Settings. Want history on a second machine? Turn on cloud sync. Don't want to? Skip it.

git clone https://github.com/deskwand/agent.git
cd agent && npm install && npm run dev

---

Install

Platform	Package
macOS	.dmg · Apple silicon
Windows	.exe · installer
Linux	.AppImage · .deb

All releases → · deskwand.com/#download

Quick start

1. Download and install
2. Add an API key from any supported provider
3. Pick a model
4. Set a goal — DeskWand plans, executes, and keeps going on its own

Documentation

• User manual — the complete guide, from first run to advanced features
• Issues — bug reports and feature requests
• AGENTS.md — how this repository is put together

Development

npm run dev            # development build
npx vitest run         # test suite
npm run build          # package for the current platform
npx eslint <files> --fix && npx prettier --check <files>

Commits follow Conventional Commits (English messages, enforced by commitlint + husky). Bug reports and pull requests are welcome.

www.deskwand.com · X