Open-source alternatives to Devin, Cursor and Claude Code

OpenHands

The All-Hands-AI team rebranded the OpenDevin project last year and the main repo now sits at 72,607 stars under MIT. The architecture is unchanged: an autonomous agent that plans, edits, runs tests, browses the web, and opens PRs from inside a sandboxed Docker container. The hosted free tier defaults to a Minimax model; the published numbers people actually quote come from runs against Claude Sonnet 4.5, where SWE-Bench Verified lands around 77%.

Who self-hosts it. Teams with hard data residency rules, engineers who want to read the agent's own planner before trusting it with a credit card's worth of API calls, and anyone with enough scale that per-seat pricing has become the budget conversation rather than a footnote.

Where it falls short. Setup is real DevOps work — sandboxing, container limits, network egress rules — and getting any of those wrong is how you end up with an agent that hangs or burns through context. A single SWE-bench-style fix costs $0.50–$3 against a frontier model, so what you save on the wrapper you'll spend on inference. The exchange is worthwhile at scale; below about 20 active developers, the math is harder.

Aider

Aider is a terminal pair programmer that edits your local git repo and auto-commits every change. The repo map is built on tree-sitter, the diffs are surgical, and every edit becomes a real git commit you can review or revert with the tools your team already uses for code review. That last property is why Aider sticks where it sticks — the audit trail isn't a proprietary log file, it's git log.

It targets Claude 3.7 Sonnet, DeepSeek R1 and Chat V3, the OpenAI o-series and GPT-4o cleanly, and connects to most of the rest via LiteLLM. Local models work if you're willing to live with the latency.

Where it falls short. Aider only edits files. It can't run shell commands, install packages, or execute tests on its own. For a fire-and-forget loop you need OpenHands. For a precise, reviewable tool that respects your git history, nothing in this list is closer to what a careful senior engineer would write themselves.

SWE-agent

SWE-agent is the academic project that put open source on the SWE-bench leaderboard, with the entire harness configured in a single YAML file. The smaller mini-SWE-agent variant lands above 74% on Verified in roughly a hundred lines of Python — the kind of result that makes the codebase itself worth reading once before you commit to building on top of anything heavier. Default model is bring-your-own; Claude, GPT and anything LiteLLM speaks all work.

Best for. Running headless against a queue of GitHub issues. Benchmarking. Building your own agent on top of a clean baseline whose decisions are all visible.

Weak for. Day-to-day "help me with this file" coding. There's no chat polish, no IDE integration, no onboarding wizard. The documentation expects you to read the source. That's a feature for one audience and a wall for another.

Continue and Cline

Cursor's pitch is a VS Code fork with AI baked in. The open-source answer is the inverse: keep stock VS Code, add the same primitives via extensions. Continue (~31k stars) handles autocomplete and chat across VS Code and JetBrains and is model-agnostic against Claude, GPT-4o or a local Ollama target. Cline (~58k stars across its Roo Code and Kilo Code forks) is the agentic extension closest to Cursor's agent mode, with the Plan/Act split that makes it noticeably safer to leave running while you're in another window.

Where they fall short. The editor itself is stock VS Code. You don't get Cursor's tab-completion model or the same end-to-end latency. If those specific UX wins are what you currently pay Cursor for, no extension closes that gap — you're trading editor polish for model freedom and a lower bill, and that's the whole exchange.

Tabby

Tabby is the only tool here designed first for an air-gapped team server rather than a single laptop. It ships as a Docker container with admin panel, LDAP auth, per-user API keys, and usage analytics built in — the things you'd otherwise have to bolt onto OpenHands yourself.

Defaults are open-weight coders: StarCoder, Qwen2.5-Coder, DeepSeek Coder. A single A100 80GB running Qwen2.5-Coder 32B at 4-bit quantization will serve roughly 15–25 concurrent developers using Tabby's request queueing — the exact number depends on completion length and how aggressive your users are with chat.

Replaces. GitHub Copilot for completion, plus a basic chat panel.

Doesn't replace. Cursor's agent mode or Devin-style autonomy. Tabby is autocomplete plus chat, served well, behind your firewall. It's not an engineer.