Wove

Open-source AI coding agent. Desktop app. Bring your own model.

AI coding tools today lock you into one model, one vendor, one monthly subscription. Your code goes through someone else's servers. You can't switch models when a cheaper one fits. You can't work offline. And you can't see what the tool actually does with your code.

Wove fixes this. It's a desktop app that connects directly to any AI provider with your own API key. It writes code, runs commands, browses the web, and verifies its own work — all locally on your machine. No proxy, no cloud agent, no vendor lock-in.

Use Claude, GPT, Gemini, MiniMax, Ollama, or any OpenAI-compatible model. Switch between them in one click. Pay only for what you use.

Built on Wave Terminal (Go + Electron). Apache 2.0.

Download

macOS (Apple Silicon / M1+): Wove-darwin-arm64-0.2.0.dmg · zip

macOS (Intel): Wove-darwin-x64-0.2.0.dmg · zip

Linux and Windows builds coming soon.

All releases

Why Wove

You own the whole stack. Wove is a desktop app that talks directly to AI providers. Your code stays on your machine. Your API keys stay in an encrypted local keystore. There's no proxy, no telemetry requirement, no usage cap.

It actually does the work. Wove doesn't just suggest code — it reads your project, writes files, runs tests, opens a browser to check the result, and fixes what's broken. When a task is too big, it splits it into sub-tasks with isolated context so nothing overflows.

Any model, one interface. Claude Opus for hard problems. GPT-5 Mini for quick tasks. MiniMax M2.7 for $3/month coding. Ollama for fully offline work. Switch in one click, same tools and guardrails everywhere.

What it does

Writes and edits code

Reads your project conventions (WAVE.md, CLAUDE.md), finds similar files for style reference, writes code that matches your codebase. Enforces read-before-edit at tool level — the AI physically cannot modify a file it hasn't read.

Runs and verifies

Executes commands in a real terminal. Runs your test suite. Checks linter output. Won't say "done" without running at least one verification step. If large edits are needed, breaks them into stages instead of rewriting entire files.

Browses the web

Built-in browser with CDP-based vision. The agent opens pages, reads content, clicks buttons, fills forms, takes screenshots with numbered element markers. Useful for testing web apps, scraping docs, or debugging frontends.

Plans and delegates

Creates execution plans with concrete file paths, tracks progress in a live panel. When tasks are complex (5+ files, multi-step migrations), automatically delegates to sub-tasks running in isolated tabs with fresh context.

Connects to your data

Auto-detects MCP servers from .mcp.json. The AI gets direct access to your database schema, documentation, logs — any MCP-compatible data source.

Models