XDA ran a piece this week under the title “my $0 local coding agent does more than $20 Claude Code because it never says no.” The point is not that a local model is smarter (it isn’t); it is that a tool you own, that runs against a model you chose, that you can pause and rewind at will, ships something usable in cases where a hosted agent refuses. In 2026 the free, local end of the AI coding tool market has quietly caught up. These are the eight desktop apps we run when we want the agent to stay on our own machine.
What to look for in a free local coding agent
Local models beat cloud ones only on availability, not raw quality. The tool matters more than the model. Pick for these:
- Model portability. Ollama, LM Studio, or a raw endpoint. If the tool locks you into one runtime, walk away.
- Real edit mode. The agent has to write diffs, not print snippets you paste. Git integration is the practical test.
- Human-in-the-loop by default. Approve, deny, edit, then run. Auto-accept modes are a footgun for local use.
- A useful context model. Something that reads @-mentioned files, tracks the current diff, and drops old messages when the window fills. Otherwise you burn tokens on the wrong thing.
- Runs where the code lives. VS Code extension, JetBrains plugin, or a terminal that respects your git repo. No shipping code to a hosted panel.
- Real logs. You can see what the agent did, and undo any part of it. Anything less is faith-based programming.
Quick comparison
| App | Best for | Free plan | Standout feature |
|---|---|---|---|
| Aider | Terminal-first git-native agent | Fully free (open source) | Automatic commits per change |
| Continue | VS Code + JetBrains with any model | Fully free (open source) | One config, different models per task |
| Cline | Autonomous VS Code coding | Fully free (open source) | Full agentic loop with approval prompts |
| Tabby | Self-hosted server for teams | Fully free (open source) | Air-gapped inference for whole orgs |
| Roo Code | Cline fork for BYOK power users | Fully free (open source) | Multiple modes, prompt engineering-first |
| OpenHands | Web-based general coding agent | Fully free (open source) | Browser + shell + editor in one loop |
| Ollama | The runtime the tools above talk to | Fully free (open source) | One-command local LLM serving |
| Kilo Code | Roo/Cline fork with a focus on stability | Fully free (open source) | Model management panel, orchestrator mode |
The 8 apps
1. Aider — best terminal-first, git-native agent
Aider is the pick if the shell is your home. Point it at a repo, name a few files, describe the change, and Aider makes edits and commits each one under its own git message. The default model behaviour is edit-mode, so you get diffs, not chat. Work with Ollama, run against DeepSeek Coder V2 or Codestral locally, and the loop feels closer to a very fast pair programmer than to a chatbot.
Where it falls short: No IDE integration. The commit granularity is opinionated and takes getting used to. Full-repo context needs manual @-mentions.
Pricing:
- Free: fully open source
- Paid: none
Platforms: Windows, macOS, Linux
Download: Aider install
Bottom line: The pick if you already do most work in a terminal. Nothing else on this list has as tight a git-plus-model loop.
2. Continue — best cross-IDE assistant
Continue is the closest thing the open-source world has to a “just install the extension and it works” AI coding assistant. It ships in the VS Code marketplace and the JetBrains marketplace from the same config file, so you can standardise a team on one setup across editors. Route different tasks to different models (Ollama for edits, a hosted model for chat, a small model for autocomplete), and the plugin handles the fan-out.
Where it falls short: The out-of-the-box configuration is safe but tame. Getting the best out of Continue means reading the config docs and tuning models per role.
Pricing:
- Free: fully open source
- Paid: Continue Hub for team-shared configs, optional
Platforms: VS Code and JetBrains on Windows, macOS, Linux
Download: Continue install
Bottom line: The pick when the team runs a mix of VS Code and JetBrains and you want one AI coding config.
3. Cline — best autonomous VS Code agent
Cline is the pick if you want the agentic loop, not just the completions. Give it a task, and it plans, edits, runs commands, and asks for approval at each step. The human-in-the-loop model is the reason it belongs on this list rather than the “walks off with your codebase” cluster. Bring-your-own-key with Ollama, together with a decent local model, gets you a real agent without a subscription.
Where it falls short: Small local models get lost on multi-step plans; you want at least a 32B-class coding model to feel the difference. UI is functional, not polished.
Pricing:
- Free: fully open source
- Paid: none
Platforms: VS Code on Windows, macOS, Linux
Download: Cline on the VS Code Marketplace
Bottom line: The pick when you want a real agent inside VS Code and are willing to run a real model.
4. Tabby — best self-hosted server for teams
Tabby is the option built for the org where “code cannot leave the building” is a hard constraint. Deploy the Tabby server on a machine with a GPU (or a CPU with the smaller models), and every developer’s IDE plugin talks to it. Everything (completions, chat, agent runs) stays on your hardware. Air-gap-friendly deployments are a first-class use case.
Where it falls short: Requires actual infrastructure. Not a “install and run” tool for a single developer.
Pricing:
- Free: fully open source
- Paid: Tabby Enterprise for support and admin features
Platforms: Linux server, IDE plugins for VS Code, JetBrains, and Vim/Neovim
Download: Tabby install
Bottom line: The pick if you are deploying an AI coding tool across a team and cannot use hosted models.
5. Roo Code — best for BYOK power users
Roo Code started as a Cline fork and evolved into a distinct tool aimed at heavy prompt-engineering workflows. Multiple modes let you switch between Code, Architect, Ask, Debug, and custom flows without restarting the session; the prompt library is first-class; and the diff review is easier than in vanilla Cline.
Where it falls short: Similar hardware demands as Cline. The mode-switching model requires you to actually structure how you work; casual use leaves half the tool unused.
Pricing:
- Free: fully open source
- Paid: none
Platforms: VS Code on Windows, macOS, Linux
Download: Roo Code on the VS Code Marketplace
Bottom line: The pick if you already use Cline and want more control over prompts and modes.
6. OpenHands — best web-based general agent
OpenHands (formerly OpenDevin) puts a full sandboxed dev environment behind a web UI. Browser, terminal, and file editor are all available to the agent; you point it at a task and watch it work in a container. Run it locally with Docker plus Ollama, and every action stays on your machine.
Where it falls short: Docker-heavy setup. The sandbox model is powerful and also easy to misconfigure into a stuck state.
Pricing:
- Free: fully open source
- Paid: OpenHands Cloud for hosted use, optional
Platforms: Docker on Windows, macOS, Linux
Download: OpenHands install
Bottom line: The pick when you want a headless dev environment for the agent to work inside, without giving it your whole laptop.
7. Ollama — best runtime for everything above
Ollama is not a coding agent; it is the runtime the other tools on this list talk to. One command pulls a model (ollama pull codestral:latest), a second command starts serving it on localhost:11434, and every editor plugin that speaks the Ollama protocol finds it. In 2026, Codestral, DeepSeek Coder V2, Qwen2.5-Coder, and Llama 3.2-Coder are all live-loaded from Ollama in seconds.
Where it falls short: Not a UI. Model selection and hardware sizing are on you; a 7B model on a laptop CPU is a very different experience from a 34B model on a workstation GPU.
Pricing:
- Free: fully open source
- Paid: none
Platforms: Windows, macOS, Linux
Download: Ollama download
Bottom line: Install this first. Every other tool on the list needs it.
8. Kilo Code — best Cline-family fork focused on stability
Kilo Code is another fork of the Cline/Roo family, and its pitch is polish over feature count. The model management panel gives you a real place to compare and switch runtimes; the orchestrator mode chains tasks for you; and the diff review handles large multi-file changes without going sideways.
Where it falls short: As with Cline and Roo, small local models struggle on multi-step plans. Some features are still catching up with upstream.
Pricing:
- Free: fully open source
- Paid: none
Platforms: VS Code on Windows, macOS, Linux
Download: Kilo Code on the VS Code Marketplace
Bottom line: The pick if Cline felt raw and Roo felt busy. Kilo splits the difference.
How to pick the right combination
Every setup here layers a model runtime under an agent front-end. Start there:
- The tightest local stack: Ollama + Aider + Codestral. Terminal, git-native, works on modest hardware.
- The tightest IDE stack: Ollama + Continue + a mixed model config (edits on Codestral, chat on Qwen2.5-Coder).
- The full autonomous loop: Ollama + Cline (or Kilo Code) + a 32B-class model. Assume a GPU on the machine.
- The org-wide answer: Tabby on a shared GPU server, with the IDE plugins on every developer machine.
- The sandboxed loop: OpenHands with Ollama. When you want the agent to run commands in a container, not on your laptop.
The point of the XDA piece is fairness: paid cloud coding agents refuse tasks that local agents complete. Even when they refuse for good reasons, the local option is worth having. Even when the paid one wins on quality, the local one keeps working when the network drops.
FAQ
Which is the best local coding model in 2026? Codestral from Mistral and DeepSeek Coder V2 are the most cited. Qwen2.5-Coder-32B lands close on many benchmarks and runs on less VRAM. Test on your codebase before committing.
What hardware do I actually need? For 7B models, most Apple Silicon Macs and any 8GB+ VRAM GPU. For 32B models, aim for 24GB+ VRAM or an M-series Mac with 32GB+ unified memory. CPU-only works but you will feel every token.
Can I combine local and cloud models? Yes. Continue and Cline both let you route different tasks to different models. Local for private code, cloud for hard reasoning tasks.
Is Aider still worth it if I use Cline? Yes. They solve different problems. Aider is best when you want tight, git-committed edits from the terminal. Cline is best for multi-step tasks with a plan.
Do any of these work offline? Every tool on this list runs offline when paired with a local model runtime. The exception is model updates, which need connectivity.
