Wuwei AI

A free, open-source, local-first AI agent client

Like Claude Code or Cursor — but free, local, open-source, and works with any model. Tell it what you want in one sentence and it reads/writes files, makes precise edits, runs commands, and searches the web to finish the task — every step behind a permission prompt.
Switch between Claude / OpenAI / Chinese LLMs in one click, bring your own key or use Wuwei's hosted credits.

Website · Download · 简体中文

_{⭐ If Wuwei saves you the setup / API-key / VPN pain, a star really helps others find it.}

_{Live demo · click to watch the HD mp4}

What is this

Wuwei is an AI agent — not autocomplete, not a chat box. At its core: a large language model + a tool-execution loop + an interface (terminal TUI / desktop GUI).

You state a goal and it reads files, edits code, runs commands, and searches the web to finish the task from start to end — while the harness is fully self-built, open source and auditable, with your code and data staying on your own machine.

🆓 Genuinely free — works out of the box, no subscription, no credit card
💻 Local-first — your project and code stay on your machine, never uploaded somewhere you can't see
🔍 Open source (MIT) — auditable line by line, no hidden telemetry, no region tagging
🔌 Bring any model — Claude / any OpenAI-compatible endpoint / local models (vLLM, Ollama) / Chinese LLMs, one-click switch
🌏 Direct access in China — no VPN needed for Chinese models
🛡️ Permission prompts — confirm before writing files or running commands (y / N / a = always allow this tool)
🧠 Brain (local knowledge network) — a local concept knowledge network that captures project / deployment / gotcha knowledge; recall subgraphs on demand via brain_recall, saving tokens instead of re-reading whole docs every time

Brain · local knowledge network

A concept knowledge network that runs locally: it structurally captures project background, git paths, test / production environments, deploy-script locations, gotchas and more, then recalls them on demand — instead of stuffing the whole doc into context every time.

brain_recall — recall the relevant concept subgraph per task (+ hits from your doc library, returning only summary + path)
brain_learn / brain_link — remember high-value knowledge and wire up relations; same-name overwrite corrects stale info
brain_read_doc — read the full text by path when you need it, no full scan required

Note: Brain is a Wuwei hosted / member capability. You can toggle it and view / override its prompt in the "Knowledge Network" settings. When off, the brain_* tools and their instructions are disabled together.

Architecture

Interface (TUI / GUI) → self-built agent main loop → tool execution (with permission prompts) ↔ model backend.
Models connect two ways: bring-your-own-key direct or the Wuwei hosted gateway.

flowchart TB
    User(["You · one-sentence goal"])

    subgraph UI["Interface"]
      TUI["Terminal TUI · Ink"]
      GUI["Desktop GUI · Electron"]
    end

    subgraph Core["Agent core · self-built harness (MIT)"]
      Loop["Agent main loop<br/>token counting · auto context compaction"]
      Perm{"Permission prompt<br/>write file / run command"}
      Tools["Tools<br/>read · write · edit · bash · glob · grep"]
      Brain[("Brain<br/>local knowledge network<br/>recall · learn · link")]
    end

    subgraph Backend["Model backends"]
      P1["Claude<br/>api-key / subscription OAuth"]
      P2["OpenAI-compatible<br/>DeepSeek·GLM·Kimi·local vLLM/Ollama"]
      P3["Codex<br/>ChatGPT subscription"]
    end

    User --> UI --> Loop
    Loop --> Perm -->|after confirm| Tools
    Tools -->|feed result back| Loop
    Loop <-->|brain_recall / learn<br/>recall on demand, save tokens| Brain
    Loop <-->|BYOK: direct| Backend
    Loop <-->|Wuwei hosted: platform credits| GW["Wuwei gateway<br/>wuweiai.io/api/gateway"]
    GW <--> Backend

Two ways to use it

Bring your own key (BYOK) — fill in your own API key or local endpoint; fully free, data never goes through a third party.
Wuwei hosted — use platform credits with zero config; even works free without login (see the website).

Quick start

Desktop (recommended)

Download the installer for your OS (Windows / macOS / Linux) at wuweiai.io and you're ready to go.

CLI / run from source

git clone https://github.com/wuwei-io/wuwei.git
cd wuwei
npm install

# Option 1: Claude API
export ANTHROPIC_API_KEY=sk-ant-...
export WUWEI_MODEL=claude-sonnet-5        # optional
npm run dev

# Option 2: local / OpenAI-compatible endpoint (vLLM / Ollama, etc.)
export WUWEI_BASE_URL=http://localhost:8000/v1
export WUWEI_MODEL=qwen3-coder
npm run dev

After it starts, just type your request. Common commands: /reset clears the conversation · /exit quits. File-write / command-run operations ask for confirmation.

Desktop dev / packaging

npm run desktop:dev          # dev
npm run desktop:build        # build
npm run pack:wuwei           # Windows installer
npm run desktop:pack         # macOS installer

Supported model backends

anthropic (API key) · anthropic + OAuth (Claude subscription / Claude Code) · openai-compatible (DeepSeek / Zhipu GLM / Kimi / MiniMax / Doubao / Qwen / Hunyuan / Grok / local vLLM·Ollama) · codex (ChatGPT subscription). Credentials are saved in separate slots per platform; switch provider / model from the bottom bar in one click.

Project structure

src/                CLI (terminal TUI)
  index.tsx           entry: builds the Agent and renders the Ink UI
  config.ts           picks the model backend from env vars
  agent/
    loop.ts           agent main loop + token counting + auto context compaction
    provider.ts       multi-backend: anthropic / openai-compatible / codex
    prompt.ts         system prompt
  tools/index.ts      tools: read / write / edit / bash / glob / grep
  ui/                 Ink TUI components
desktop/            desktop (Electron)
  main/               main process: provider presets, account credits, key vault
  renderer/           renderer: multi-session, streaming, Markdown, images

Saving tokens

Long tasks won't break: when the context exceeds WUWEI_COMPACT_THRESHOLD (default 80% of the window) old history is auto-compacted into a summary; WUWEI_KEEP_RECENT (default 6) keeps the most recent turns.

Contributing

Issues and PRs welcome. Run npm run typecheck before submitting to make sure types pass.

Star & support

Wuwei is built in the open by a tiny team. If it's useful to you, starring the repo ⭐ is the single biggest thing you can do to help it reach more people — it takes one click and genuinely makes a difference. Bug reports and PRs are just as welcome.

wuwei