From DeepReinforce, Ornith hands task writing, scaffolding and solution writing to the model itself, comes in three sizes from 9B to 397B, and ships under MIT with a 256K context and local deployment.

An open source desktop companion that sends several agents into their own branch worktrees, meters the context each one burns, and keeps the diff and the editor in the same window.

A community continuation of Roo Code that turns one agent into four fixed roles, shows every change as a diff before it lands, and throws a guard over dangerous commands during long unattended runs.

The network is off by default and writes stop at the working directory, with edits and commands passing an approval step; rewritten in Rust it is one static binary that also runs unattended in a pipeline.

An open version of a well-known online tool, rebuilt so the model is yours to choose: code runs inside a Node.js sandbox in the browser tab, dependencies install, a development server starts, and the result appears in a preview beside the conversation. Output arrives as discrete steps, so a failure can be repaired at the step that broke.

A research agent that takes a single bug report and carries the repair to the end: it clones the project, reads the description, hunts for the relevant files, edits them, runs the tests and hands back a patch that can be merged. Its real contribution is not the model but the interface built for reading and editing code.

An open desktop builder: describe the application you want, and it generates a conventional web project on your own disk, runs it on a local port and refreshes as you talk. Every round is versioned and can be rolled back, the code stays yours in a clean repository, and a hosted backend can be attached when the thing needs to store data.

An open coding agent that closes the loop by itself: it writes the change, runs the tests, reads the failures, finds the cause and repairs it, mostly without being asked. Bigger tasks are split across sub agents working in separate branches at the same time, and the model behind the work can be swapped at any point.

An open coding agent that lives in the command line: it loads the entire repository into context, takes a request in plain language, locates the files, makes the edit, runs the tests and corrects itself until they pass. Planning mode, parallel helper agents, reusable skill packages, and a model layer you can point anywhere.

An assistant you host yourself as a single service process, with no external database and no cloud dependency. Completion, in-editor questions and an administrative console are packaged together, the repository is indexed so answers refer to your own code, and a mainstream graphics card is enough to serve a small team.

An AI pair programmer that lives in the terminal, so the editor never has to be opened. It reads the repository, proposes a change, applies it, runs the tests, and commits every modification to version control with a written message, leaving a clean path back. Free and open source; only model calls cost money.

An open coding agent that lives in the editor. It reads the project, proposes a plan, and only then touches files; every change leaves a checkpoint you can revert. You supply the model key, so nothing is marked up and no repository leaves your machine. Also available as a command line tool and a library.

An open platform for software engineering agents. Each session runs inside its own container, which is destroyed afterwards, so installing dependencies and running risky commands cannot touch the host. Any model can be attached, and the same agent is reachable from a terminal, a browser, an SDK or a hosted service.

A deliberately small coding agent for the terminal: four tools, a system prompt under a thousand tokens, and everything else left to an extension system the agent can write for itself. No built-in planning mode, no task list, no bundled tool servers, and a clear account of why each was left out. Free and open source.

A terminal coding agent in a single statically linked binary, designed around one constraint: keeping the beginning of a conversation byte-for-byte stable so cached input stays cheap. Long sessions, a plan to approve before anything is written, per-turn checkpoints and automatic repair of malformed tool calls. Free and open source.