Open-Source AI Coding Assistants in 2026: Cline vs Aider vs Continue vs OpenHands

Updated September 2026: rewritten with each project’s current status, and with trial evidence on speed and security. Continue is no longer maintained.

Open-source AI coding assistants promise the best of both worlds. You get an AI pair programmer, but you keep control over the code, the model and the data.

Four names come up again and again: Cline, Aider, Continue and OpenHands.

A lot has changed in 2026. One of the four has stopped development altogether, and another has gone quiet.

The research has moved on too. The most careful trial so far found that AI tools made experienced developers slower, not faster.

Here’s where each tool stands today, and what the evidence says about using any of them.

Which open-source AI coding assistant should you use?

The short answer depends on where you work and how much autonomy you want to hand over.

That’s a guide to fit, not a ranking. None of these tools has been tested head to head in an independent trial.

The rest of this article explains what each one does, based on the projects’ own documentation, and what research says about the risks.

What changed in 2026

Continue reached the end of the road

The biggest change is Continue. Its GitHub page now carries a clear note: the repository “is no longer actively maintained and is read-only for all users.”

The team says it “did a final 2.0.0 release” of its VS Code extension, command-line tool and JetBrains plugin. That release removed anonymous telemetry and authentication.

The code remains open source under the Apache 2.0 licence. So others can build on it, but you shouldn’t expect fixes or new features.

Aider has gone quiet

Aider is still available and widely used. But its GitHub repository was last updated on 22 May 2026.

Its latest tagged release, version 0.86.0, came out in August 2025. For a tool in such a fast-moving field, that’s a long pause.

That doesn’t mean it has stopped working. It does mean you should check its activity before relying on it for a team.

Cline and OpenHands kept growing

Cline has spread well beyond its original VS Code extension. It now offers a command-line tool, a desktop app for macOS and Windows, a JetBrains plugin and a developer kit.

OpenHands now calls itself “the self-hosted developer control center for coding agents and automations”. It can run agents locally, in Docker containers, on virtual machines or on company servers.

The four tools at a glance

Here’s each project as of 30 September 2026, from GitHub.

Stars measure popularity, not quality. They’re a rough sign of how many people might help fix problems.

Cline: the IDE choice

What it does

Cline works inside your editor. According to its documentation, it can create files, run terminal commands, browse the web and use tools.

It has a Plan mode, where it explores your code and proposes a strategy. Then an Act mode carries the plan out.

The safety feature that matters

Cline’s documentation says “every file edit and terminal command requires your approval”. That keeps a human in the loop.

It also works with many AI providers, including Anthropic, OpenAI and Google. You can run local models too, through tools like Ollama or LM Studio.

Who it suits

Teams who live in VS Code or JetBrains, and who want to review each change. Project-specific rules can be stored in `.clinerules` files.

Aider: the terminal choice

What it does

Aider describes itself as “AI pair programming in your terminal”. It can connect to “almost any LLM, including local models”.

The git advantage

Its standout feature is version control. Aider “automatically commits changes with sensible commit messages”.

That means you can use normal git tools to see, review and undo what the AI did.

Who it suits

Developers comfortable in the terminal, who want every AI change tracked. Just keep an eye on how actively it’s maintained.

Continue: what to do if you use it

Your options

If Continue is already part of your setup, it won’t stop working overnight. But without maintenance, security fixes and compatibility updates won’t arrive.

For a team, that’s a reason to plan a move. The most direct alternatives inside an editor are Cline, or a hosted assistant your organisation already approves.

The upside of open source

Because the code is Apache-licensed, anyone can fork it. The team says it hopes the codebase “continues to serve as a foundation for others.”

OpenHands: the agent platform

What it does

OpenHands is less a pair programmer and more a platform for running coding agents. Its documentation describes starting conversations and automating everyday engineering tasks.

It works with “any LLM”, and can also run other agents, such as Claude Code, Codex or Gemini.

The warning to read first

OpenHands can run with or without a sandbox. Its own documentation warns that without one, “the agent will have full access to your filesystem!”

A Docker sandbox keeps the agent more contained. For anything beyond experiments, that’s the safer route.

Who it suits

Teams that want agents doing longer tasks on their own infrastructure, and who have the skills to secure them.

Do AI coding assistants make you faster?

A surprising trial

This is the question that matters most, and the answer is less obvious than the marketing suggests.

In 2025, the research group METR ran a randomised controlled trial with 16 experienced open-source developers. They completed 246 real tasks on mature projects they knew well.

Each task was randomly assigned to allow or ban AI tools. With AI allowed, developers mostly used Cursor Pro with Claude 3.5 and 3.7 Sonnet.

In the researchers’ words, “AI tooling slowed developers down.” Experts in economics and machine learning had predicted the opposite.

Why this matters for tool choice

The trial didn’t test Cline, Aider, Continue or OpenHands. But it shows that feeling faster and being faster are not the same thing.

Its developers had years of experience on their own projects. Results could differ for newer developers, unfamiliar code or later AI models.

The security catch

AI can make code less secure

A large user study, first published in 2022, tested people solving security-related coding tasks. Those with an AI assistant “wrote significantly less secure code” than those without.

Worse, people with AI help were more likely to believe their code was secure.

The study also found a useful pattern. Participants who trusted the AI less, and worked harder on their prompts, produced code with fewer vulnerabilities.

Why self-hosting doesn’t fix this

Running an assistant yourself gives you control over where your code goes. It doesn’t make the code the AI writes any safer.

That still needs code review, tests and security checks, exactly as it would for human-written code.

What earlier research found

The early speed-up

A 2023 controlled experiment with GitHub Copilot asked developers to build a simple web server in JavaScript. Those with the AI finished 55.8% faster.

That’s a very different result from the METR trial. The task was small and self-contained, rather than real work on a large, familiar codebase.

The first security warning

A 2021 study of GitHub Copilot’s code prompted it with 89 scenarios involving well-known security weaknesses. Of the 1,689 programs it produced, about 40% were vulnerable.

How it fits together

AI assistants can speed up small, well-defined tasks. On complex work in code people know well, the gains may vanish.

Security risks have shown up repeatedly. So human review remains essential, whichever tool you choose.

How much should you trust this?

Early. The facts about each tool come from the projects themselves. But independent research on these specific tools is almost nonexistent.

What makes it convincing

• Project status, licences and features come directly from each tool’s GitHub page and documentation.
• The METR study was a randomised trial on real tasks.
• Security problems appeared in both a code analysis and a user study.

What makes me cautious

• No independent trial has compared these four tools with each other.
• The METR trial had only 16 developers, using early-2025 tools.
• The security studies used older AI models, which have since improved.
• Project activity can change quickly, in either direction.

What this means for you

• Check the project is alive. Look at recent commits and releases before building a workflow around any tool.
• If you use Continue, plan a move. It will keep working for now, but it won’t get fixes.
• Keep approvals on. Tools like Cline ask before edits and commands. Don’t switch that off for real projects.
• Use a sandbox for agents. OpenHands warns that without one, its agent can access your whole filesystem.
• Measure, don’t guess. Developers in the METR trial felt faster while actually being slower. Track real outcomes.
• Review AI code like any other code. Tests and security checks still matter.

The METR study page on arXiv is short and readable, and worth sharing with your team.

In this talk, Joel Becker, the study’s first author, explains how METR measured developer productivity:

What we still don’t know

1. Which open-source tool works best? No independent head-to-head trial exists.
2. Do newer models change the speed result? The METR trial used early-2025 tools.
3. Are agents safe to run unattended? There’s little research on real-world agent failures.
4. Will Aider pick up again? Its activity has slowed, but it isn’t archived.
5. Who will carry Continue forward? A community fork may appear, or may not.

My take: control is not the same as productivity

What I like about open-source assistants is the control they offer. You choose the model, you decide what runs, and your code can stay on your machines.

I’m less convinced by the promise that any of them will simply make a team faster. The best trial so far points the other way for experienced developers.

So I’d pick an actively maintained tool, keep a human approving changes, and measure whether it actually helps. Right now, that means Cline for editor work and OpenHands for sandboxed agents.