Best Open-Source AI Coding Agents in 2026: OpenCode, Codex, Pi and More

Short version: the terminal coding agent went from novelty to default in eighteen months, and the biggest change of 2026 is who ships them. OpenAI's Codex CLI, Google's Gemini CLI and Mistral's Vibe are all open source now, while Anthropic's Claude Code, the tool most people picture when they hear "coding agent", is not. Underneath them a second wave of harnesses and runtimes has appeared that treats the agent as something you orchestrate rather than talk to. This list ranks the open-source agents we would actually install, with star counts, licences and release dates read from the GitHub API on 10 November 2026, plus the projects that quietly stopped. It is refreshed monthly.

1. anomalyco/opencode (225,767 stars)

OpenCode is the most-starred coding agent on GitHub and is not backed by a model vendor. It started life as sst/opencode and now lives under the anomalyco organisation; do not confuse it with opencode-ai/opencode, the Go project that briefly shared the name, was archived in September 2025 and continued as Charm's Crush (number 10 below). It is MIT licensed, written in TypeScript, works with any model provider, installs with one curl command or npm i -g @opencode/cli, and now has a desktop app in beta. v1.19.5 shipped on 28 October. If you want one agent that is not tied to one lab's pricing, start here.

2. openai/codex (134,638 stars)

OpenAI open-sourced the Codex CLI in April 2025 and has since rewritten it in Rust; rust-v0.167.0 landed on 6 November. It runs locally in your terminal, installs into VS Code, Cursor and Windsurf, has a desktop app (codex app), and sits alongside Codex Web, OpenAI's cloud agent, for work you want done while you sleep. Apache-2.0 licensed, and it can also drive local open-weight models through codex --oss. The obvious choice if your organisation already pays OpenAI.

3. earendil-works/pi (121,166 stars)

Pi is the surprise of the year: created in August 2025, past 120,000 stars by November 2026. It describes itself as a minimal, extensible agent harness: a unified LLM API, an agent loop, a terminal UI and a coding agent CLI that you customise with extensions, skills, prompt templates and themes, bundled as Pi packages. v1.5.1 shipped on 9 November. It is for people who want to shape the agent rather than accept a vendor's defaults, and it has spawned its own ecosystem, including oh-my-pi (38,494 stars), a variant with the IDE wired in.

4. google-gemini/gemini-cli (108,750 stars)

Google's terminal agent is Apache-2.0 and ships fast (v0.67.0 on 4 November), but read the README before installing: as of November 2026 it says the CLI is designed for Gemini Code Assist Standard and Enterprise licences and Code Assist for GitHub, that free individual access and the Google AI Pro and Ultra plans moved to the Antigravity CLI on 18 June 2026, and that it is not optimised for Gemini 4 and later. For a company on a Code Assist licence it is still the supported path; for everyone else, Antigravity is now the front door.

5. OpenHands/OpenHands (93,094 stars)

OpenHands began as OpenDevin, the open-source answer to Devin, and in 2026 it changed shape again: the README now leads with Agent Canvas, a self-hosted control centre that runs OpenHands' own agent alongside Claude Code, Codex, Gemini, Pi, OpenCode or any ACP-compatible agent across local, remote and cloud backends, with automations and a Slack integration. v1.38.0 shipped on 10 November, MIT licensed. Pick it when you want to run several agents on your own infrastructure under one roof.

6. cline/cline (71,482 stars)

Cline is the agent that made VS Code users comfortable with autonomy, and it now runs in three places: the IDE extension, a terminal CLI (npm i -g cline) that can run fully headless for CI and scripting, and a native desktop app for macOS and Windows that schedules routines. Apache-2.0, v4.1.26 on 26 October. The SDK is the quiet strength: the same agent can be embedded in your own tooling.

7. openinterpreter/openinterpreter (68,850 stars)

Open Interpreter reinvented itself as a coding agent optimised for low-cost and open-weight models, and its README says plainly that the new codebase is a fork of OpenAI's Codex, which is why the interface feels familiar. The target is explicit: Kimi K3 and GLM, with a Rust reimplementation of the provider-recommended harness. rust-v0.0.60 shipped on 9 November; the original Python project lives on separately at endolith/open-interpreter. If your constraint is tokens per dollar rather than peak intelligence, this is the one to benchmark.

8. aaif-goose/goose (56,418 stars)

Block's goose is a general-purpose agent that happens to be excellent at code: desktop app, CLI and API, any LLM, extensible through MCP. It is Apache-2.0 and now lives under the aaif-goose organisation because it was one of the founding projects of the Agentic AI Foundation in December 2025, alongside MCP and AGENTS.md. v1.58.1 shipped on 10 November. Goose is the pick for teams that want one agent for code and for everything around it.

9. QwenLM/qwen-code (29,153 stars)

Alibaba's Qwen Code is open source inside and out: the framework and the Qwen models it is tuned for are both open. It runs in the terminal, editor, desktop, browser and chat, and ships Auto Memory, Auto Skills, sub-agents, agent teams and MCP support out of the box. The CLI shipped v0.27.1 on 9 November. The strongest option if you want an end-to-end open stack, provided you pair it with open-weight models.

10. charmbracelet/crush (29,029 stars)

From the people behind Bubble Tea and Glow, Crush is the best-looking terminal agent in this list and the most flexible about models: pick from a wide range or add your own OpenAI- or Anthropic-compatible endpoint, and switch models mid-session without losing context. v0.100.0 shipped on 5 November. One caveat: it is licensed under FSL-1.1-MIT, a source-available licence that converts to MIT after two years, so read the terms before using it commercially.

Also worth a look

• Kilo-Org/kilocode (27,924 stars, MIT, v7.8.27): an all-in-one agentic engineering platform that descends from Roo Code and Cline, with a CLI its README describes as a fork of OpenCode. Note the README: main no longer accepts new features, which now go to the v2 branch.
• github/copilot-cli (11,392 stars, v1.0.108 on 9 November): the same agentic harness as GitHub's Copilot coding agent, in your terminal. The repository is public but holds only the installer, changelog and a custom licence that permits installing, running and redistributing unmodified copies; the source itself is not published.
• mistralai/mistral-vibe (5,356 stars, Apache-2.0, v3.1.1 on 9 November): Mistral's minimal CLI coding assistant, the European option.
• SWE-agent/SWE-agent (20,771 stars): the research agent that takes a GitHub issue and tries to fix it. Its own README now recommends the simpler mini-swe-agent going forward.

The harness wave

The fastest-growing repos around coding agents in late 2026 are not agents at all. They wrap, route and host the agents above:

• stablyai/orca (about 120,000 stars): a workspace for running a fleet of coding agents side by side.
• ruvnet/ruflo (81,390 stars, v3.75.0): an agent meta-harness for Claude Code and Codex that turns one task into a coordinated swarm; npx ruflo@latest init and you are in.
• herdrdev/herdr (49,864 stars, created March 2026): the runtime your coding agents live on; sessions keep running when you close the client or lose SSH, and after a machine restart it restores the layout and can resume supported agent sessions, though the original processes do not survive.
• musistudio/claude-code-router (38,075 stars): started as a router for Claude Code and now positions itself as one local gateway for every agent (Claude Code, Codex, OpenCode, Pi, Kilo and others) across models and providers.
• esengine/DeepSeek-Reasonix (36,148 stars, created April 2026): a newer agent, not a DeepSeek product despite the name, built for long, complex engineering tasks; too young for the top ten, growing fast enough to watch.

Not open source, still the elephant in the room

Anthropic's claude-code repository has 155,146 stars and ships a release almost daily (v2.1.322 on 10 November), but GitHub lists no open-source licence for it: the repository tracks issues and releases for a proprietary tool. That is not a criticism, just the fact that makes everything above interesting. Most of the harnesses in the previous section exist precisely to put an open layer around Claude Code and Codex.

Faded in 2026

• Aider-AI/aider: still 50,045 stars and still a fine tool, but the last release (v0.86.0) was August 2025 and the last push May 2026. Use it, do not build on it.
• continuedev/continue: shipped a final 2.0.0 in June 2026; the README now says the repository is no longer actively maintained and is read-only.
• RooCodeInc/Roo-Code: archived May 2026 after the extension shut down on 15 May; its README points users to ZooCode, a community fork, and back to Cline. Kilo Code also descends from it.
• voideditor/void, smallcloudai/refact and microsoft/vscode-copilot-chat: all archived in the first half of 2026.
• plandex-ai/plandex: last push October 2025.

How to choose

• You pay one lab already: Codex (OpenAI) or Gemini CLI (Google Cloud). Claude Code if that lab is Anthropic, accepting that it is closed.
• You want provider independence: OpenCode or Crush.
• You want to shape the agent yourself: Pi.
• You run agents for a team: OpenHands' Agent Canvas, with herdr or orca underneath if you run many at once.
• Cost is the constraint: Open Interpreter with an open-weight model, or Qwen Code for a fully open stack.
• You live in VS Code: Cline.

How this list is built

Stars, licences, languages, release tags and archive status come from the GitHub API, captured on 10 November 2026; positioning statements are taken from each project's own README on the same day. To rank in the top ten a project must be an installable coding agent (harnesses and routers are listed separately), carry a licence GitHub recognises or at least publish its terms, and have shipped a release in 2026. We re-run the numbers monthly and log every change at the top of this page. For the frameworks these agents are built on, see our AI agent repositories list, and for the tool layer they all share, the best MCP servers.