Best Open Source Coding Agents in 2026
Open source here means a public repo and a named license. Pick by surface (terminal, IDE, desktop), model routing, and whether “done” can be the same session that wrote the patch.
Open source coding agents
License, surface, then the done contract
critique.sh
There is no single best open source coding agent in 2026. Use OpenCode for a MIT-licensed, provider-agnostic terminal (and desktop/IDE) agent. Use Aider for git-native pair programming in the terminal (Apache 2.0). Use Cline when you want an Apache 2.0 agent in the IDE and terminal with human-in-the-loop tool approval. Use Goose when you want Block’s Apache 2.0 local agent (desktop, CLI, API) under AAIF stewardship. Use OpenHands (MIT) when you want a self-hosted control center plus an OSS agent SDK. Use Gemini CLI when you want Google’s Apache 2.0 terminal agent aimed at Gemini. Continue is still Apache 2.0 source, but the public repo is read-only after a final 2.0.0 release. CritiqueCode is Critique’s author agent (@critiquedotsh/harness, binary critique-code). This post does not call it open source. None of these replace a human merge decision.
“Best open source coding agent” is a search query that mixes three different jobs: inspectable tooling, a surface you will actually live in, and a definition of finished. The first is a license plus a public repo. The second is terminal vs IDE vs desktop. The third is whether the same loop that edits files also gets to declare the change safe.
This guide only lists projects with a public GitHub repository and a named license observed on 16 August 2026. Star counts are snapshots, not quality scores. If a claim is not on the project’s own docs or LICENSE file, it is omitted. For the broader (including proprietary) field, see Best coding agent in 2026. For Claude Code specifically, see Best Claude Code alternatives in 2026.
What is an open source coding agent
A coding agent is not autocomplete. Autocomplete predicts tokens in the buffer you already have open. A coding agent takes a task, searches a working directory, edits files, runs commands, and iterates until it stops or you stop it.
An open source coding agent is that loop in a repository you can clone. OpenCode’s GitHub README calls itself “the open source AI coding agent” and is MIT-licensed at anomalyco/opencode. Aider’s FAQ states the tool is Apache 2.0 on Aider-AI/aider. Cline’s README says Apache 2.0 © 2026 Cline Bot Inc. Goose’s GitHub license is Apache 2.0. Continue’s LICENSE is Apache 2.0. OpenHands is MIT. Gemini CLI is Apache 2.0.
That is the floor. It is not a ranking. A MIT TUI and an Apache 2.0 VS Code extension can both be “open source coding agents” and still be the wrong product for your week.
How to choose an open source AI coding agent
Start with the job, not the star count. Frontier models rotate. The surface, permission prompts, and git habits stay in the team for months.
- LICENSE file and whether GitHub still marks the repo as the license you think it is.
- Surface: terminal TUI, VS Code / JetBrains extension, desktop app, or SDK.
- Who approves file writes, shell, browser, and network.
- Bring-your-own provider vs a first-party model default.
- Whether maintenance is active. Read-only archives are still source, not a roadmap.
- Whether “I ran the tests” is allowed to count as independent review.
- Undated “#1 OSS agent” listicles with no LICENSE links.
- Star counts without a retrieval date.
- Claims that local models make review unnecessary.
- Self-review comments from the same agent that wrote the patch.
- Treating a frozen extension as a living vendor product.
Open source coding agents compared
| License (GitHub) | Surface | Models | Best when | |
|---|---|---|---|---|
| OpenCode | MIT (anomalyco/opencode) | Terminal TUI; desktop app and IDE extensions documented on opencode.ai | Provider-agnostic. Project docs push OpenCode Zen as optional curated models; Claude, OpenAI, Google, and local models are in the README pitch. | You want a Claude Code-shaped loop you can inspect and point at many providers. |
| Aider | Apache 2.0 (Aider-AI/aider) | Terminal pair programmer; editor watch mode documented on aider.chat | Cloud and local LLMs via API keys / LiteLLM-style routing described in project docs. | You want git commits as the unit of work and a repo map, not a full desktop platform. |
| Cline | Apache 2.0 (cline/cline) | VS Code extension, JetBrains plugin, CLI, and SDK per the README | Bring your own provider. Optional hosted “Cline” API is a separate product surface with its own terms. | You want IDE-native agent tools with explicit approval of edits and commands. |
| Continue | Apache 2.0 (continuedev/continue) | VS Code extension, JetBrains plugin, CLI (final 2.0.0 release) | Historically BYO models in-editor. Treat current hosting/auth as frozen-release behavior. | You need Apache 2.0 editor-agent source you can still run, knowing the upstream repo is read-only. |
| Goose | Apache 2.0 (block/goose / AAIF goose) | Desktop app, CLI, and API. README: native agent for code and other local workflows. | Multi-provider, including local. Block’s original announcement and GitHub both describe any-LLM routing. | You want a local-first general agent with Linux Foundation / AAIF stewardship, not only a coding TUI. |
| OpenHands | MIT (OpenHands/OpenHands) | Agent Canvas control center, CLI, SDK. Self-hosting docs on the project. | Own agent plus ACP-compatible third-party agents documented on openhands.dev. | You want a self-hosted developer control center around coding agents, not only a laptop TUI. |
| Gemini CLI | Apache 2.0 (google-gemini/gemini-cli) | Terminal agent (@google/gemini-cli) | Gemini via Google login or API / Vertex-style auth documented in the project. | You want a first-party Google CLI you can inspect, not a multi-provider clone of OpenCode. |
| CritiqueCode | Not classified as open source in this post | Author CLI: @critiquedotsh/harness, binary critique-code | Critique Inference critique/auto after login, with optional BYOK in settings | You want implement, then forced independent review and verified repair. See docs, not a LICENSE claim. |
OpenCode: MIT terminal agent, many providers
OpenCode is the project most people mean when they say “open source Claude Code.” The public repo anomalyco/opencode is MIT-licensed. GitHub reported on the order of 197k stars on 16 August 2026. The README’s differentiators versus a lab CLI are explicit: 100% open source, not coupled to one provider, TUI-first, built by neovim users.
Official docs cover agents and a config reference. A desktop app is distributed from GitHub releases and opencode.ai/download. OpenCode Zen is the project’s optional curated model path. You can still point the same UI at Claude, OpenAI, Google, or local models. That is the product bet: models commoditize, the harness should not be owned by one lab.
Choose OpenCode when inspectable tooling and provider choice are the requirement. Do not treat star velocity as a SWE-bench score. OpenCode does not, in its public docs, force a separate reviewer after the author claims a unit of work. Head-to-head with Critique’s author agent: CritiqueCode vs OpenCode.
Aider: Apache 2.0 pair programming in git
Aider is older than the 2025-2026 CLI agent wave and still a serious option. Aider-AI/aider is Apache 2.0. GitHub reported on the order of 48k stars on 16 August 2026. The official site describes AI pair programming in your terminal: a tree-sitter repo map, automatic commits with descriptive messages, lint and test after edits, and cloud plus local LLMs.
Aider’s own FAQ states Aider AI LLC publishes the tool under Apache 2.0 on GitHub. That is the sourced company fact. The useful product fact is smaller: every AI edit is a normal git commit you can diff, revert, or review with tools you already have. Architect-style two-model flows (planner vs editor) exist in the docs as a cost control, not as independent review.
Pick Aider when the unit of work should be a commit, not a long TUI session you cannot unwind. Skip it if you wanted a desktop agent OS. Aider is a lean loop. That is the point.
Cline: Apache 2.0 agent in the IDE and terminal
Cline publishes cline/cline under Apache 2.0. GitHub reported on the order of 66k stars on 16 August 2026. The README’s job statement is “the open source coding agent in your IDE and terminal”: VS Code extension, JetBrains plugin, CLI, and an SDK (@cline/sdk).
The sourced workflow is human-in-the-loop. Docs and README emphasize creating files, running commands, browsing, and requiring approval. That is a permission model, not a second reviewer. Plan vs Act is a Cline product split for how aggressive the same agent is allowed to be.
Cline Bot Inc. has stated in a public GitHub issue that the Apache 2.0 license covers the extension when you bring your own keys, and that cline.bot/tos applies if you choose their hosted API provider. If compliance is why you wanted open source, that boundary is the thing to verify, not a slogan on a roundup.
Continue: Apache 2.0 source, read-only upstream
Continue still ships as Apache 2.0 on continuedev/continue. GitHub reported on the order of 35k stars on 16 August 2026. The README calls it a pioneering open-source coding agent for CLI, VS Code, and JetBrains.
The same README states the repository is no longer actively maintained and is read-only, after a final 2.0.0 release that removed anonymous telemetry and authentication and squashed bugs. Treat Continue as inspectable, frozen software you can still run, not as a vendor with an active OSS roadmap on that repo. If you need a living IDE agent with a public license, Cline is the closer 2026 option on the facts above.
Goose: Apache 2.0 local agent from Block / AAIF
Goose is an extensible agent that runs on your machine. GitHub lists Apache 2.0 on block/goose (also referenced as aaif-goose/goose). Stars were on the order of 53k on 16 August 2026. Block’s Open Source Program Office announced “codename goose” as Apache License 2.0, connecting LLMs to real-world actions, with software engineering as the first use case.
The current README describes a native desktop app (macOS, Linux, Windows), a CLI, and an API, built in Rust. It is explicit that the agent is not only for code: research, writing, automation, data analysis. MCP extensions are part of the public docs. Stewardship language around the Agentic AI Foundation / Linux Foundation appears on project and community pages. For this list, the sourced license is Apache 2.0 on GitHub. Ignore third-party pages that still say MIT if the repo LICENSE says Apache 2.0.
Pick Goose when you want a local-first general agent with a public license, not when you only wanted a coding TUI with the tightest edit loop. Aider and OpenCode are narrower. Goose is broader on purpose.
OpenHands: MIT agent platform you can self-host
OpenHands is MIT-licensed at OpenHands/OpenHands (also known as All-Hands-AI/OpenHands). GitHub reported on the order of 84k stars on 16 August 2026. The site describes an open source platform for software engineering agents: GUI, CLI, SDK, and an enterprise control plane. The README currently leads with Agent Canvas, a self-hosted control center that can run the OpenHands agent or ACP-compatible third-party agents.
The Software Agent SDK (OpenHands/software-agent-sdk) is also MIT. That is a different shape from a laptop TUI: you are buying into a control plane and optional self-hosting, not only a pair-programming CLI. Commercial OpenHands Cloud exists next to the MIT code. This post does not treat hosted SKUs as the open-source agent. The MIT repos are.
Gemini CLI: Google’s Apache 2.0 terminal agent
Gemini CLI is Google’s open-source terminal agent (Apache 2.0), package @google/gemini-cli, docs at geminicli.com. The README’s job is the shortest path from a prompt to Gemini, including coding, file work, and command execution.
It belongs on an open-source list because the CLI is Apache 2.0. It does not belong in the “model-agnostic” column. Auth and quotas live on Google’s docs and change. Compare it to OpenCode only if your actual need is Gemini in a first-party CLI you can read.
CritiqueCode: author agent, not listed as open source here
CritiqueCode is Critique’s author agent. You install @critiquedotsh/harness and run critique-code in the repository you intend to edit. The session implements. After the author claims the unit complete, the controller forces independent review and verified repair. Author transcripts are not Evidence. That split is the product. Docs: CritiqueCode. Comparison index: comparisons, including CritiqueCode vs OpenCode.
This post does not call CritiqueCode open source. There is no LICENSE claim in this roundup. If you need a public MIT or Apache 2.0 harness, use one of the projects above. If you need an author CLI that cannot skip review, that is a different job. It is also not the Critique CLI sidecar (@critiquedotsh/cli, binary critique). Do not install the wrong package.
Published package
npm install --global @critiquedotsh/harness
cd /path/to/your-project
critique-codeBinary is critique-code. Package name is @critiquedotsh/harness.
How to pick without a fake winner
OpenCode if the constraint is a MIT, provider-agnostic terminal (or desktop) agent with a large public community.
Aider if the constraint is git-native edits, automatic commits, and a thin Apache 2.0 CLI.
Cline if the constraint is an IDE-resident Apache 2.0 agent with explicit tool approval, plus a CLI/SDK if you grow into them.
Goose if the constraint is a local-first Apache 2.0 agent (desktop + CLI + API) with Block / AAIF provenance, including non-code workflows.
OpenHands if the constraint is a MIT control center and SDK you can self-host, not only a pair-programmer in tmux.
Gemini CLI if the constraint is Google’s models in a first-party Apache 2.0 terminal agent.
Continue only if you accept a read-only Apache 2.0 snapshot. CritiqueCode if the constraint is the completion contract, not the license label. More comparisons live at /comparisons.