An open source AI coding assistant gives you the agent's code, not the AI. Every tool still needs a model behind it: an API key for a hosted model, or a local model on your own hardware. If you connect a hosted model, your code goes to that provider, so decide where the model runs before you choose the tool.
The most common mistake with open source coding assistants is assuming the open licence covers the whole thing. It covers the agent: the part that reads your files, plans changes and edits code. The thinking is done by a language model, and that model is a separate decision with its own terms.
Below are nine tools, in no particular order, starting with a short table of what form each takes and its licence. After that comes the question that matters most, then the tools grouped by where your code and the model run.
Form and licence at a glance
| Tool | Form | Licence |
|---|---|---|
| Aider | Terminal pair programmer | Apache-2.0 |
| Cline | Editor extension and CLI agent | Apache-2.0 |
| Codex CLI | Terminal coding agent | Apache-2.0 |
| goose | Desktop and terminal agent | Apache 2.0 |
| gptme | Terminal agent, local-first | MIT |
| mini-swe-agent | Minimal research-grade agent | MIT |
| OpenHands | Self-hosted agent platform | MIT, self-hosting guide linked |
| Qwen Code | Terminal coding agent | Apache-2.0 |
| Tabby | Self-hosted completion and chat | Apache 2.0 outside its ee/ directory |
Nearly all of these use a permissive licence, Apache or MIT, so you can run and change them with few conditions. The exception to watch is an open core project: the self-hosted completion server in the table is Apache 2.0 outside its ee/ directory, and the code inside that folder is under a separate licence. A source-available licence isn't the same thing either, which is why Charm's Crush isn't on this page.
The model behind the agent is your call, and it decides privacy
Open source doesn't make the AI free to run or private. Each of these agents sends your prompts, and the code it reads, to whichever model you connect. That leaves two routes:
- A hosted model through an API key. Easy to set up and usually the strongest results, but your code leaves your machine and goes to that provider under its terms.
- A local model on your own hardware. Your code stays with you, but you install, run and update the model yourself, and a laptop may struggle with larger ones.
- A server you host for the team. Some projects document self-hosting so a whole team can share one setup inside your own network. You maintain it.
Aider, gptme and mini-swe-agent document local models. Tabby and OpenHands document self-hosting. For the rest, model choice is the vendor's description, so test the one you want before you commit. And remember that a few well-known open source coding tools have been archived or folded into other products, so a repo's recent activity is worth checking as of Nov 2026.
Where your code and the model run
In your editor, or on a server you host
Cline sits in your editor or on the command line. The completion server and the agent platform with its control canvas are the two to look at if you want one setup the whole team uses on your own infrastructure. Both mean someone looks after a server.
In your terminal, pointed at the model you choose
Qwen Code, Codex CLI, goose and the two local-first terminal tools mentioned above work from the command line. The OpenAI agent's repo is Apache-2.0 and runs locally, but its README also points to separate cloud and desktop products from the same company, so don't assume every feature you've seen elsewhere is in the open repo.
A minimal agent for software issues
The smallest tool here is deliberately bare. It's built for working through software issues from the command line and suits developers who want to see every step the agent takes.
All nine, in no particular order
Aider
Best for: Pair programming with a language model in your terminal, on a new or existing codebase, with a map of the whole project built as you work.
Cline
Best for: Developers wanting an autonomous agent inside their editor, on the command line or as an SDK, with open model choice as the vendor describes it.
Codex CLI
Best for: Running a coding agent from OpenAI locally in your terminal, separate from the company's cloud and desktop products.
goose
Best for: Teams wanting an extensible agent that installs, edits, runs and tests code, plus a development kit for building agents into their own apps.
gptme
Best for: Terminal users who want a personal agent that works over ssh and in CI, and handles general tasks beyond coding, per its README.
mini-swe-agent
Best for: Solving software issues from the command line with a deliberately small agent, local model support documented, from the team behind SWE-agent.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
OpenHands
Best for: Running several coding agents from one self-hosted control centre, which its README says can drive other ACP-compatible agents too.
Qwen Code
Best for: Working from the terminal with the option of VS Code, Zed and JetBrains integrations listed in its README.
Tabby
Best for: Code completion, inline chat and an answer engine in your IDE, served from infrastructure you run so requests can stay in-house (vendor claim).
My advice before you pick one
Start with the model, not the agent. If your code can't leave your machine, choose a tool that documents local models or self-hosting, and check your hardware can run a model you're happy with. If a hosted model is fine, pick the agent that fits how you already work: the editor if you live in it, the terminal if that's home.
Try two at most on a small, low-risk repo, and compare the changes they make rather than how they feel to use. If you're weighing which model to run locally, our guide to open source language models covers that side. For broader agents that handle non-coding work, see our list of AI agents for small businesses.
Before an agent edits your repo
- Work on a branch. Let the agent change a copy, never your main line of code.
- Read every diff. These tools can write code quickly, and they can also write wrong code quickly.
- Keep secrets out of reach. Check which files the agent can read before connecting a hosted model, especially config and environment files.
- Limit what it can run. Agents that run commands and tests should ask before anything destructive.
- Check the licence of the model too. A local model has its own terms, separate from the agent's licence.
Do you make an open source AI coding assistant or agent? Tool makers can ask to be considered for a featured place on this list.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
Frequently asked questions
Is an open source AI coding assistant free?
The agent's code is free to download as of Nov 2026, but the model behind it isn't always. You either connect a hosted model through an API key or run a local model on your own hardware.
Does an open source coding agent keep my code private?
Only if the model runs on hardware you control. If you connect a hosted model, the code the agent reads is sent to that provider.
Can I run an AI coding assistant completely offline?
Some tools here document local models, which lets you work without sending code anywhere. You'll install and run the model yourself, and larger models need strong hardware.
What does open core mean for a coding assistant?
Most of the code is under an open licence, but part of it sits under separate terms. One tool here keeps an ee/ directory under a different licence from the rest.
Is an editor extension or a terminal agent better?
Neither is better in general. An extension suits people who work in their editor all day; a terminal agent suits people comfortable on the command line and in scripts.
Can a team share one self-hosted coding assistant?
Yes, a few projects document self-hosting for that. Someone on the team will need to run the server and keep it updated.
