I open-sourced my paid MCP server. 40 tools, Apache 2.0, runs locally on Mac/Windows.

So I've been running openowl as a paid product for a few months. $5/mo for unlimited tool calls, signed binary, hosted dashboard. 200 signups, 14 paid.

Last week I ripped out the licensing code and pushed the core as Apache 2.0.

It's a Python MCP server that gives any AI assistant (Claude, Codex, anything MCP-compatible) eyes and hands on your desktop. The 40 tools cover screenshots, clicking, typing, scrolling, OCR, accessibility tree queries on macOS, pywinauto on Windows, workflow recording and replay, and window watchers that fire when something changes on screen.

Install (cloning works today, PyPI publish queued):

git clone https://github.com/mihir-kanzari...
cd openowl
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[macos]"

Then point Claude Desktop at it:

{"mcpServers": {"owl": {"command": "/path/to/.venv/bin/openowl"}}}

Why open-source when it was making money? The funnel was leaky. Six percent of signups ever made their first tool call, and most of the people who never came back said the install was the friction, not the price.

The moat was never the code, it was the signed binary and the hosted account flow. I'm keeping both of those for the cloud version ($5/mo unlimited at openowl.dev), but opening the core means anyone who wants to audit it, fork it, or self-host can do that without paying me anything.

It's alpha. Coordinate clicking on dense UIs misses sometimes. The vision-based finder is more reliable but slower. Workflow replay assumes the screen looks roughly like when you recorded, which is a strong assumption.

Roast it, fork it, or tell me what's missing.

Repo: github.com/mihir-kanzariya/openowl