Coding agents you can run on your own machine

Most coding agents default to a cloud model, but several open-source ones run happily against a model on your own machine through Ollama, LM Studio or vLLM. That matters when the code can’t leave the building, or when you just want to understand what the agent is doing.

A quick map, from the tools’ own docs and comparisons (MarkTechPost, OpenHands):

Cline. Lives in your editor. Every file edit and command asks for approval by default, which is the right default. The VS Code extension, CLI and SDK are Apache-2.0. The JetBrains plugins aren’t open source.
goose. Started at Block and moved to the Linux Foundation’s Agentic AI Foundation in April 2026. Its strength is MCP: dozens of documented extensions, plus local inference through Ollama.
OpenHands. Built for longer, more autonomous tasks. It has the most detailed documentation for running local models with LM Studio, Ollama, vLLM or SGLang.
Codex CLI. Apache-2.0, with built-in support for Ollama and LM Studio. The catch is that it speaks only the Responses API, so your local server has to expose that endpoint.

The shared limitation

Tool calling. Local models handle simple, well-defined tool use fine, but on long chains of steps they still fall behind frontier cloud models. In practice that means local agents shine at bounded tasks (refactor this file, write tests for this function) and struggle with “go build the feature”.

Two habits worth keeping

Sandbox it. Some agents have no permission system at all and their own docs recommend running them in Docker. Even with approvals, an agent that can run shell commands should not run on the machine with your SSH keys.
Read the licence. “Open” models are often open-weight under custom licences, not open source in the strict sense. Check before you ship something built on one.

This space changes monthly, so treat the list as a snapshot from October 2026, not a ranking.