CursorBuddy: an open, multi‑platform agent
Jason Kneen released CursorBuddy, an open‑source MIT agent for Cursor that runs on desktop, CLI, web and voice and supports plugins—positioning it as a lightweight, community‑driven alternative for multi‑modal developer assistants. The project highlights continued interest in portable, extensible agents outside closed commercial stacks (x.com/jasonkneen/status/2042211288785338408)
A lot of coding assistants live inside one editor window. CursorBuddy is trying the opposite: one little agent can run as a browser embed, a desktop overlay, or a command line tool, and the same project docs say it can also speak, listen, and take screenshots across multiple displays. (github.com 1) (github.com 2) The basic idea is simple enough to picture. Instead of a chat box pinned to a sidebar, CursorBuddy shows up as a blue triangle that follows your mouse, flies to buttons and coordinates, and changes shape when it is idle, listening, processing, or responding. (github.com) That matters because modern coding work now happens in more than one place. Cursor itself now has a command line interface for macOS, Linux, and Windows, so an assistant that wants to stay useful has to move between the editor, the terminal, and whatever is happening on screen. (cursor.com 1) (cursor.com 2) Jason Kneen’s repository is already laid out around that portability. The project includes a web build for script-tag embeds, an Electron desktop app for a system-wide overlay, and a command line package under `packages/cli`, which means the same “buddy” is being treated more like a reusable layer than a single app window. (github.com 1) (github.com 2) The desktop version is the clearest example of what he is aiming at. The docs say `npm run dev:electron` launches a transparent overlay window that follows the real system cursor plus a separate panel window from the tray icon, so the assistant can point at things happening anywhere on the operating system instead of only inside one tab. (github.com) The voice piece is not a bolt-on microphone button. CursorBuddy’s docs describe a full voice pipeline with push-to-talk input, speech-to-text from AssemblyAI, Deepgram, or OpenAI Whisper, and text-to-speech from ElevenLabs or Cartesia, which turns the assistant into something closer to a talking copilot than a text-only bot. (github.com) (github.com) It also tries to stay model-agnostic. The inference docs list Anthropic, OpenAI, Ollama, and LM Studio, so a developer can point the same interface at cloud models or local models instead of being locked to one provider’s stack. (github.com) The plugin story is where this gets more interesting. CursorBuddy can load custom tool files from project folders or from `~/.cursorbuddy/tools/`, hot-reload them when they change, and merge those tools with Model Context Protocol servers before sending them to the model. (github.com) Model Context Protocol is the plumbing that lets one agent call tools exposed by another process, like giving a contractor a standard wall socket instead of hard-wiring every appliance. CursorBuddy’s docs say it works both ways: it can connect to outside Model Context Protocol servers, and it can expose itself as a server so other agents can control the cursor, capture screenshots, or trigger speech. (github.com) That puts it in a different lane from the usual “one company, one assistant, one interface” approach. The docs explicitly name Claude Code and Codex as clients that can control CursorBuddy through Model Context Protocol, which makes CursorBuddy look less like a rival model and more like a portable body that other agents can borrow. (github.com) One correction to the early chatter around the launch: the current GitHub repository is not MIT-licensed. The repository pages and docs both label CursorBuddy as `AGPL-3.0-only`, which is a strong copyleft license that requires networked derivatives to share source, so the “open” part is real but the license is more restrictive than a permissive MIT release. (github.com) (github.com) The bigger signal here is that developers still want assistants they can move around, rewire, and inspect. When a project ships with browser embedding, desktop overlay mode, command line packaging, model switching, filesystem tools, and Model Context Protocol hooks in the same repo, it is betting that the next useful agent will behave less like a closed app and more like a portable layer that can sit on top of whatever tools a developer already uses. (github.com) (github.com)