A minimal yet ready-to-use AI agent harness, built on best practices for frontier models.

Web app · Documentation · HTTP API · 简体中文

---

Wish is a minimal, fast AI agent that runs on your own machine. Use it from the web app on desktop or phone, or over the HTTP API.

 

Why Wish

• Minimal and fast. One Rust program with its own embedded database; nothing else to install or run.
• Tools and skills load on demand. MCP servers and skills (the SKILL.md folders other agents use too) are looked up and called from the shell only when a task needs them, so adding more never grows the model's context or resets its prompt cache.
• Lean tool design. A few general tools: a shell with background jobs and exact file edits, image viewing, web search, questions to you and a search of its own history. Everything else goes through the shell.
• Ready to use. Install it, run wish and open the browser. Presets cover the major model providers and coding plans, ChatGPT and GitHub Copilot subscriptions, the Magpie gateway and local models.

Quick start

1. Install Wish with whichever package manager you use:

npm install -g wish-agent               # Linux, macOS and Windows
yay -S wish-agent-bin                   # Arch Linux
brew install windusth/tap/wish-agent    # macOS and Linux

The packages install the wish command, which can't sit beside Tk's wish on Arch Linux or Homebrew.

2. Start it.

wish

3. Open http://127.0.0.1:8790. A short first-run setup helps you add a model provider. Pick your working directory on the start page and send your first message.

The first start writes a configuration for your user (~/.config/wish-agent/config.json on Linux) and keeps its data beside it. See configuration for every option, and deployment for running Wish as a service and reaching it from other devices.

From source

You need a Rust toolchain (stable) and Node.js 22.19 or newer.

git clone https://github.com/WindustH/wish-web.git
(cd wish-web && ./pnpmw install --frozen-lockfile && ./pnpmw build)
git clone https://github.com/WindustH/wish-core.git
cd wish-core
cargo build --release
cp -r ../wish-web/dist target/release/web
./target/release/wish

Without the web app

Everything the web app does is available over the HTTP API:

# Create a session with shell access in /tmp
curl -s http://127.0.0.1:8790/api/sessions -H 'Content-Type: application/json' -d '{
  "provider": "openai", "cwd": "/tmp", "tools": {"shell": true},
  "config": {"model": "gpt-5", "stream": true, "tools": [], "run": {"tools": "Serial"}}}'

# Send it a message; it starts working right away
curl -s http://127.0.0.1:8790/api/sessions/SESSION_ID/input \
  -H 'Content-Type: application/json' -d '{"text": "What is in this directory?"}'

# Follow along live
curl -N http://127.0.0.1:8790/api/sessions/SESSION_ID/events

Documentation

Configuration	Every option in config.json, providers and presets
Deployment	Running as a service, access control, remote access, backups and upgrades
HTTP API	Endpoints, event streams and error handling
Internals	How the engine works, for contributors

Security

Wish is built for one trusted person. Anyone who can reach its API can run commands with the permissions of the account Wish runs under. By default it listens only on 127.0.0.1, and without a token it answers only requests addressed to it by its own name, so a web page cannot reach it through your browser. Before exposing it, set an access token and put it behind HTTPS; deployment explains how.

Contributing

Issues and pull requests are welcome. Wish is written in Rust (edition 2024); cargo build is all it takes. The test suite lives in a separate wish-test repository and exercises the real binary over HTTP against a local mock provider, so no API keys or network access are needed.

License

MIT