Open source AI agents

Open source AI agents are coding assistants whose source code you can read, run and fork. As of 31 October 2026 the active ones are OpenCode (MIT), Cline, Goose, Codex CLI, Gemini CLI and Zoo Code (all Apache-2.0), OpenHands and OpenClaw (MIT), plus Continue and Aider. Roo Code is shut down and Flowise is archived. The software is free; the model is not.

The agents, side by side

Every figure below comes from the project's GitHub repository on 31 October 2026. Stars are a rough signal of attention, not of quality.

Two rows deserve a second look. Roo Code is archived: the README states "The Roo Code Extension was shut down on May 15th." Aider still has a large user base, but its last tagged release is from August 2025 and the last merge on the main branch we saw was on 22 May 2026. That is not dead, and it is not busy either.

OpenAI's and Google's agents are open source in the licensing sense, yet each is designed for its maker's models. That distinction matters later in this page.

What "open source" means here

Three different things get called open source in this market, and mixing them up costs people time.

1. Open code. You can read the repository, build it and change it. Everything in the table above passes this test, because each repository carries an OSI-style license (Apache-2.0 or MIT) in its LICENSE file.
2. Open models. The model weights are published, so you can run them yourself. This is a separate question from the agent, and it is covered in open vs closed models and the list of open source LLMs.
3. Open governance. Who decides what goes in. Goose now sits under the Linux Foundation's Agentic AI Foundation, and its README says so. OpenClaw is run by a foundation. Most others are controlled by a single company or team.

An agent can be open code and still send your files to a closed model. Open code does not by itself keep your data on your machine. For that you need an open model running locally, covered in run an LLM locally and Ollama.

Agent by agent

OpenCode

An MIT-licensed terminal, desktop and IDE agent with the largest star count among the coding-first tools. It reaches 75+ providers through Models.dev and can also sign in with a Copilot or ChatGPT account. Read the full OpenCode guide for install lines and the Zen service. It is the closest open source answer to a general-purpose Claude Code, and the Claude Code vs OpenCode page compares the two.

Cline

Cline describes itself as "the open source coding agent in your IDE, terminal, & desktop". Its README lists a CLI (npm i -g cline), a desktop app for macOS and Windows, a VS Code extension, a JetBrains plugin and an SDK (npm install @cline/sdk). One detail from the repository index: the JetBrains plugin is not open source, since the README says "Currently we are not open-sourcing JetBrains plugins". Everything else in the index points to code in the repo. Plan and Act modes, .clinerules files and MCP support are all documented. See Cline.

Goose

Goose is a general-purpose agent, not only a code tool. The README describes "a native desktop app for macOS, Linux, and Windows", a CLI and an API, built in Rust, working with 15+ providers and 70+ MCP extensions. It can also use existing Claude, ChatGPT or Gemini subscriptions through ACP. The repository moved from the Block organisation to aaif-goose, and the docs moved to goose-docs.ai, so older links may redirect. See Goose.

OpenHands

OpenHands changed shape. The repository README now leads with "Agent Canvas", a self-hosted control center that runs the OpenHands agent and also drives Claude Code, Codex, Gemini CLI, Pi and OpenCode through the Agent Client Protocol. It installs with npm install -g @openhands/agent-canvas. Its README warns that the no-sandbox option gives the agent full access to your filesystem. See OpenHands.

Codex CLI and Gemini CLI

Both are Apache-2.0 and both are open on GitHub. Codex CLI is OpenAI's terminal agent, and Gemini CLI is Google's. They are best read as open front ends for one vendor's models. Gemini CLI has an extra caveat in 2026: it is now tied to Code Assist enterprise, and individuals moved to Antigravity CLI on 18 June 2026. Details are on Codex CLI and Gemini CLI.

OpenClaw

OpenClaw is an MIT-licensed assistant that runs on your own computer and talks to you through chat apps. Coding is one of several jobs it can do. It is the most-starred project in the table by a wide margin, but it is a different kind of tool from a terminal coding agent. See OpenClaw.

Aider

Aider is the veteran: "AI pair programming in your terminal". Install with python -m pip install aider-install then aider-install. It builds a map of your repository, edits with git commits, and works with almost any model, including local ones. The README's badges claim 6.8M installs and 15B tokens per week; those are the project's own numbers. The slowdown in releases is the thing to weigh. See Aider.

Continue

Continue is an Apache-2.0 IDE extension and CLI project. We did not write a standalone page for it. Check its repository before relying on it: the latest tagged release we saw was v2.0.0-vscode on 19 June 2026, though the repo was pushed to on 30 October.

Roo Code and Zoo Code

Roo Code started as a fork of Cline and was popular for its modes (Code, Architect, Ask, Debug and custom modes). The company wound down the extension on 15 May 2026 to focus on Roomote, according to the Zoo Code README. Zoo Code is a fork by former contributors, Apache-2.0, with a migration guide from Roo. It is young: about 2,000 stars against Roo's 24,000. If you used Roo, the choice is between Zoo Code (closest) and Cline (the original). See Roo Code.

What open source buys you

• No lock-in to one model vendor. OpenCode, Cline, Goose, Aider and Continue let you swap providers. Your habits and your rules files stay.
• Readable behaviour. When an agent does something odd, you can read the prompt and the tool code. Closed agents give you a changelog.
• Pinning. You can stay on one version while a vendor ships changes you do not want.
• Self-hosting. OpenHands is designed to run on your machine, in Docker or on a server you control. For company use, that can matter more than any feature.
• Forking as an exit. Roo Code's shutdown shows why this matters. The community forked the code and kept going. A closed product that shuts down leaves you with nothing.

What it does not buy you

• Free usage. The model bill is still there. A heavy day on a frontier model through an API key can cost more than a flat subscription, and the flat plans from closed vendors are often subsidised. See AI coding agent pricing compared.
• Privacy by default. Cloud models see your code. Read each provider's terms. OpenCode's site says it does not store your code, with an exception for its free models.
• Polish. Closed agents such as Claude Code and Cursor tend to have tighter editor integration and faster releases of vendor-specific features. Open agents catch up, but not on the same day.
• Support. There is no account manager. For a team, check whether the maintainer offers a paid plan.
• Safety. An agent that can run commands can damage your machine. Open code lets you audit it; it does not sandbox it for you. Use a container or a dev VM for risky tasks.

Costs in practice

The agents themselves cost nothing, so costs come from four places.

Using a subscription login in a third-party tool depends on the provider allowing it. Anthropic, OpenAI and GitHub have each changed such terms over time.

Local models are the only route with no per-token bill and no outside data flow. The tradeoff is quality and speed on ordinary laptops. A guide to hardware is on GPU requirements for LLMs, and local coding agent with Ollama walks through the setup.

MCP, rules files and moving between agents

All of the active agents above speak MCP, the open standard for connecting tools. The details differ: OpenCode takes an mcp key in its config, Cline has an MCP marketplace, Goose calls them extensions. Because MCP is shared, a server you set up once is cheap to move. Start with the MCP server list and best MCP servers.

Project instructions are less uniform. OpenCode reads AGENTS.md. Cline reads .clinerules. Claude Code reads CLAUDE.md. If you want portability, write one AGENTS.md and point the others at it. Our guide to writing a good AGENTS.md explains the format.

Choose by situation

If you only want to pay nothing, see free AI agents. For a wider look including closed tools, see AI coding agents.

Risks to check before you commit

1. Is it maintained? Look at the release date and the last push. Roo Code and Flowise show that a popular project can stop.
2. Who owns it? A foundation, a company or a hobbyist? A company can shift focus; Roo did.
3. Are all parts open? Cline's JetBrains plugin is closed. Some projects keep cloud features proprietary.
4. What leaves your machine? Prompts, file contents and sometimes telemetry. Read the privacy section of the README.
5. What can it run? Check the approval model. Prefer agents that ask before commands, and run them in a container.

Our take

This is an editorial reading of repositories and READMEs, not a test.

Open source is a good reason to choose an agent only if you name which benefit you want. If it is model freedom, the shortlist is OpenCode, Cline, Goose and Aider. If it is self-hosting for a team, OpenHands. If it is just a lower bill, the answer depends more on the model than the agent.

The 2026 lesson is about durability. Two well-known projects stopped in the same year, and one of them (Roo Code) was saved by a fork only because the code was open. So open source is less a guarantee that the tool lives and more a guarantee that you are not stuck if it stops. Pick one that is actively released, keep your rules in portable files, and avoid building a workflow around a single maintainer.

Checklist

• Decide what you want from open source: model freedom, self-hosting, audit or cost
• Check the latest release date and last push on GitHub
• Confirm the LICENSE file for the exact repo you install
• Choose how you will pay for the model: key, subscription or local
• Write a shared AGENTS.md so you can switch tools
• Run the agent in a container for anything touching production data
• ] Review [how we rate before treating any list as a ranking