Local-first
Runs on Ollama, vLLM or LM Studio on your own machine or home server. Your code never has to leave your network. Prefer a frontier model for a hard problem? Switch to Claude, GPT, Gemini or OpenRouter in one command.
Free & open source · by JKagiDesigns
hexiel is an AI coding agent for your terminal and browser. It reads your project, runs your tests, fixes bugs and builds features — on local models running on hardware you already own, or on any big cloud model when you want one.
AI coding tools are amazing — and most of them are a monthly bill wired to someone else's servers. hexiel gives you the same kind of agent, but you own it, you choose the model, and it's designed to be fast on the computer you already have.
Runs on Ollama, vLLM or LM Studio on your own machine or home server. Your code never has to leave your network. Prefer a frontier model for a hard problem? Switch to Claude, GPT, Gemini or OpenRouter in one command.
Anything that doesn't need a model — searching, mapping the project, waiting on tests, parsing logs — runs locally in Python. A lean prompt halved the time to process each request on an Intel iGPU mini-PC (46.5 s → 23.7 s).
Open source under GPLv3 with the Commons Clause. Use it, change it, share it, contribute back — free. It just can't be repackaged and sold without permission, so no one can turn it into the subscription it was built to replace.
A fast full-screen terminal UI and a browser UI share the same brain — plus one-shot headless runs for scripts and an OpenAI-compatible API, so tools like OpenWebUI can talk to it.
/ for a command menu with descriptions/auto
Reads your project's AGENTS.md / CLAUDE.md, keeps file-based memory, and
tracks multi-step work in a todo.md that it maintains automatically — so if your
laptop dies mid-task, hexiel -c resumes from an accurate checkpoint.
Open /config (or ⚙ Settings in the browser), click a value, pick from the real options —
including the models your server actually has — and choose this session or
save permanently.
/update and resumes your conversation
Type @shot.png and vision models see it.
pytest, vitest, Maven, Gradle, Go, Cargo, Docker, Podman.
Search + page reading built in, no API key needed.
Claude Code skills and OpenWebUI tools work unchanged.
Crashes become incidents; /heal proposes a reviewed fix.
/ship turns your work into a GitLab merge request.
One local script instead of ten slow model round-trips.
Linux, macOS and Windows (PowerShell built in).
Claude Code, opencode and Aider are excellent tools — we use and respect them. Here's an honest look at where hexiel fits. Facts checked against each project's official docs on .
| Feature | hexiel | Claude Code | opencode | Aider |
|---|---|---|---|---|
| Open source | YesGPLv3 + Commons Clause | Noproprietary | YesMIT | YesApache-2.0 |
| Price of the tool | Yesfree — pay only for cloud models you choose | PaidPro/Max/Team plan or pay-per-token API | Yesfree; optional paid model plans | Yesfree; you pay your LLM provider |
| Local models (Ollama, vLLM, LM Studio…) | Yestuned for modest hardware | NoClaude models only (officially) | Yes | Yes |
| Terminal UI | Yesfull-screen | Yes | Yes | Partialchat REPL, not full-screen |
| Browser / desktop UI | Yesbuilt-in browser UI | Yesweb, desktop and IDE apps | Yesweb UI; desktop app | Partialexperimental --gui |
| Headless / scripting | Yes | Yes | Yes | Yes |
| Project memory + auto-compaction | Yesplus auto-maintained todo.md and memories | Yes | Yes | Partialconventions file via config |
| Uses Claude Code skills | Yesreads ~/.claude/skills | Yesnative | Yesreads .claude/skills | No |
| Screenshot / image input | Yes | Yes | Yes | Yes |
| Self-update | Yes/update, keeps your conversation | Yes | Yes | Yes |
| MCP servers | Nonot yet | Yes | Yes | Nonot in official docs |
| Resale by others | Nonot without permission | No | YesMIT allows it | YesApache-2.0 allows it |
Sources: Claude Code license, Claude Code model gateways, opencode providers, opencode web, opencode skills, Aider options, Aider + Ollama. Spotted something out of date? Tell us and we'll fix it.
Measured on an Intel iGPU mini-PC, not a datacenter. Searching, mapping, waiting on tests and parsing
run in Python; a python tool batches many steps into one local script; the prompt is kept lean.
Thinking can be switched off per model for speed.
Multi-step requests become tracked tasks with a checkpoint refreshed after every step, and durable facts are saved to memory during compaction — in Python, at zero token cost.
Serve hexiel as an OpenAI-compatible model that OpenWebUI (or any OpenAI client) can connect to, and run existing OpenWebUI tool, pipe and filter plugins inside hexiel without porting them.
Crashes are recorded as incidents; /heal reproduces, fixes, tests and proposes a merge
request, and /ship turns your work into one. Nothing merges without a human.
If you live on Claude models and want deep IDE and desktop integration, Claude Code is superb. If you need MCP servers today or a permissive MIT license, opencode is a great choice. If you want a lightweight git-centric pair programmer, Aider is proven. If you want an agent that is free, runs well on modest local hardware, maintains its own memory and todo, and plugs into OpenWebUI — that's what we built hexiel for.
You need Python 3.11+ and a model: Ollama for local, or an API key for a cloud provider.
git clone https://gitlab.com/jkagidesigns1/public/applications/hexiel.git
cd hexiel
python3 -m venv .venv
.venv/bin/pip install -e '.[serve]'
ln -s "$PWD/.venv/bin/hexiel" ~/.local/bin/hexiel
ollama pull qwen3.6:35b-a3b-coding # or use a cloud key
cd ~/your-project && hexiel
winget install Microsoft.PowerShell
git clone https://gitlab.com/jkagidesigns1/public/applications/hexiel.git
cd hexiel
python -m venv .venv
.venv\Scripts\pip install -e ".[serve]"
ollama pull qwen3.6:35b-a3b-coding # or use a cloud key
.venv\Scripts\hexiel
Run hexiel --doctor any time for preflight checks with copy-paste fixes, or
hexiel --serve for the browser UI.
hexiel is designed and built by JKagiDesigns LLC, a software studio in Kingston, Massachusetts. Our day job is building custom business software for small companies — software that replaces the subscriptions that cost too much or do too little.
We wanted an AI coding agent that worked the same way: one you own. One that runs on the mini-PC in the closet instead of a monthly bill, that doesn't waste tokens watching a test suite scroll by, and that you can open up and improve. So we built hexiel — a guardian angel for your codebase — and we're giving it away.
The name nods to that idea of a watchful helper; the logo is a pair of wings around an H, under a halo.
Visit JKagiDesigns.com →License: GPLv3 with the Commons Clause. Want to build something commercial on top of it? Talk to us.
Yes. The tool costs nothing and has no subscription. If you use a cloud model you pay that provider for what you use; local models are free.
Yes — point it at Ollama or any OpenAI-compatible server on your machine or LAN. Web search obviously needs the internet, but everything else works locally.
Our test machine — an Intel mini-PC with integrated graphics — runs a 35B mixture-of-experts coding model at about 12 tokens/second. That model needs roughly 25 GB of memory, so plan for 32 GB of RAM or more. Smaller machines can run smaller models, or use a cloud model.
Mixture-of-experts models with ~3B active parameters that support tool calling and vision —
qwen3.6:35b-a3b-coding passed all our checks. Set reasoning_effort = "none" on slow
hardware so it doesn't spend minutes "thinking".
Yes. hexiel reads ~/.claude/skills directly, plus OpenWebUI tool/pipe/filter plugins.
By default every edit and command asks for your approval first. The browser UI is local-only unless you explicitly expose it, and it warns you loudly if you do.
Fork, branch, open a merge request — or let hexiel do it with /ship. Changes are reviewed
and approved by the maintainers.