The most capable open source AI coding agents in 2026 are Codex CLI (Apache-2.0), OpenCode (MIT), Cline (Apache-2.0) and Kilo Code (MIT). All four read your code, edit files and run commands, and you can read, build and change every line of them. What open source doesn’t give you is a free model: unless you run one locally, every step still goes to a paid API. Claude Code, Copilot CLI and Cursor are not open source.
Open source AI coding agents compared
| Agent | License | Interface | Models | Status |
|---|---|---|---|---|
| Codex CLI | Apache-2.0 | Terminal | OpenAI with a ChatGPT plan or key; local via Ollama or LM Studio; providers that speak the Responses API | Active, by OpenAI |
| OpenCode | MIT | Terminal, desktop app (beta), IDE extension | 75+ providers; ChatGPT and Copilot sign-in; Ollama, LM Studio, llama.cpp | Active |
| Cline | Apache-2.0 | VS Code, JetBrains, CLI, desktop app | Many providers with your keys | Active |
| Kilo Code | MIT | VS Code, JetBrains, CLI, desktop app (beta) | 500+ models via Kilo Gateway, your keys, Ollama, LM Studio | Active, owned by Anaconda since July 2026 |
| goose | Apache-2.0 | Desktop app, CLI, API | 15+ providers; Claude, ChatGPT or Gemini plans via ACP | Active, at the Linux Foundation |
| Qwen Code | Apache-2.0 | Terminal, IDE plugins, desktop, web | OpenAI, Anthropic, Gemini and Qwen APIs; Ollama, vLLM | Active, by the Qwen team |
| Gemini CLI | Apache-2.0 | Terminal | Gemini, for Code Assist Standard/Enterprise and paid API keys | Maintained for enterprise use |
| Pi | MIT | Terminal, SDK, RPC | Many providers via API keys or OAuth | Active |
| Hermes Agent | MIT | Terminal UI, chat apps, desktop | Any model: Nous Portal, OpenRouter, OpenAI, your own endpoint | Active, general-purpose |
| OpenHands | MIT | Self-hosted web app (Agent Canvas) | Any LLM; also runs other agents over ACP | Active |
| Aider | Apache-2.0 | Terminal | Most cloud and local models | Last release February 2026 |
| DeepSeek Harness | MIT | Local web UI | DeepSeek, other providers, OpenAI-compatible endpoints | Developer preview |
| Contenox | Apache-2.0 | Desktop app | ChatGPT and Claude (through Claude Code) plans, Copilot unofficially; API keys; local | Active |
Which one should I pick?
- A terminal agent close to Claude Code: Codex CLI if you pay for ChatGPT; OpenCode if you want to pick the model.
- An agent inside your editor: Cline or Kilo Code. Both run in VS Code and JetBrains with your own keys.
- Local models only: OpenCode or Kilo Code, both of which treat Ollama and LM Studio as normal providers. See run local LLMs for coding.
- A desktop app instead of a terminal: goose, the OpenCode desktop app or Contenox.
- Something you shape yourself: Pi. It ships small on purpose and skips plan mode, sub-agents and permission prompts; you add what you need.
- An open-source AI editor: Zed, licensed mainly under GPL-3.0-or-later. Void, the other well-known one, is deprecated.
What open source does and doesn’t get you
It gets you:
- Code you can audit. You can see what the agent sends, where, and which commands it is allowed to run.
- No lock-in to the tool. If the company changes direction, the code stays. When Roo Code was archived in May 2026, its code stayed available for anyone to fork.
- Model choice, usually. Most open-source agents take any provider. Codex CLI is the exception that proves the point: it is open source but built around OpenAI, and custom providers must speak OpenAI’s Responses API.
It doesn’t get you:
- A free model. The agent costs nothing; the tokens don’t. Agent work sends a lot of them: one month of the Contenox team’s own use on the DeepSeek API was 23,815 requests and 7.86 billion tokens, about 330,000 tokens per request. See AI coding costs.
- Privacy by default. Your code still goes to whichever model provider you use. Only a local model keeps it on your machine.
- No telemetry by default. Open source means you can check, not that nothing is sent. Codex sends usage metrics unless you set
analytics.enabled = falsein its config. Aider collects analytics only if you opt in. Continue removed its anonymous telemetry in its final release. Read each tool’s settings. - Safety. Pi has no permission prompts by design, and DeepSeek Harness asks you to read its safety notice before running it. Open code doesn’t mean careful defaults.
The source-available trap: some tools publish their code under licenses that aren’t open source. Crush uses the Functional Source License, which only turns into MIT two years after each release. Claude Code’s repository is public, but its license reserves all rights. Copilot CLI’s license lets you run and redistribute it only unmodified. “On GitHub” and “open source” are not the same thing.
Is Claude Code open source?
No. The Claude Code repository is public, but the code is used under Anthropic’s commercial terms. The closest open-source agents are Codex CLI and OpenCode. See Claude Code alternative.
Is IBM Bob open source?
No. IBM Bob is a commercial product, priced per instance, with an IDE, a terminal tool and an on-premises option. It routes between Claude, Mistral and IBM’s own Granite models; Granite is open (Apache-2.0), Bob is not. See IBM Bob.
Is GitHub Copilot open source?
Partly. Copilot Chat for VS Code is open source: its code moved into the main VS Code repository, under MIT. Copilot CLI is not. Either way, both need GitHub’s Copilot service, which is not open source.
Which open-source coding agents are no longer maintained?
- Roo Code: archived on 15 May 2026; read-only, no more fixes. Kilo Code and Cline are the usual next stops.
- Continue: no longer actively maintained after its final 2.0.0 release; read-only.
- Void: deprecated, no longer accepting contributions.
- Aider: not archived, but its last release on PyPI was in February 2026.
An unmaintained agent still runs, but it won’t learn new models’ tool formats or get security fixes.
Contenox
Contenox is an open-source (Apache-2.0) desktop app, built Windows-first. You open a project folder, pick a provider, including your ChatGPT or Claude plan or a local model through Ollama, and the agent works there while you approve. What it allows without asking is a set of rules you can read in Permissions, not a model’s judgment. Download it here.
Related guides
- Open source AI agents: beyond coding: goose, Hermes, Agent Zero, OpenHands and the frameworks
- Best AI coding agents: the ranking, open source or not
- What is an AI coding harness?: what the agent adds on top of the model
- AI coding harness comparison: features side by side
- Terminal coding agents: install Claude Code, Codex CLI and OpenCode
- Codex alternative: including what “open-source alternative” means for an open-source tool