Bring your own model — an embedded GGUF file, any Ollama / LM Studio / vLLM server, or Claude — and get a real agent loop with a permission engine, checkpoints, and MCP tools. Nothing has to leave your machine unless you tell it to.
Most local-model tools — Ollama, LM Studio — stop at serving a model. localagent is a full coding agent built on top of one: a real read → act → check loop, a deterministic permission engine (not LLM judgment), one-checkpoint-per-task revert, and per-hunk diff review before anything gets written.
Local-first isn't a label bolted on here. Inference for the embedded and custom Hugging Face models runs entirely in-process — zero backend, zero telemetry, by construction — not a setting you have to hunt down and switch off.
The permission engine is plain code with four modes and a command-risk classifier, so what runs without asking is a rule you can read, not a model's guess:
# no server, no other app — a GGUF model runs in-process
node dist/cli.js "explain how add() works in math.js" \
--workspace ./fixture-repo \
--mode DEFAULT
Settings, MCP servers, the model picker, a custom Hugging Face search, picking a workspace, multiple session tabs, a real task with a diff to approve, and signing in — cut together from unscripted runs of the real app.
See the real colored diff before approving an edit — not a model's summary of what it plans to do, the actual patch. Uncheck the hunks you don't want; only the checked ones get written to disk.
Beyond the 10 curated models, search Hugging Face's GGUF listings from inside localagent, or type any hf:org/repo:quant path directly — same in-process runtime as the curated models, nothing downloaded until you pick one.
Connect local MCP servers — GitHub, Postgres, anything speaking the protocol. Every call still asks for your approval, exactly like a shell command; nothing runs silently just because it came from an MCP tool.
Run up to 6 sessions at once. Each keeps running and updating in the background whether you're looking at it or not.
Every task's first write snapshots your workspace via a scratch git index — never touching your real index, HEAD, or branch. Revert restores everything, including files it created.
Ctrl+K / ⌘K to jump to any saved session or open a panel, instead of hunting through the sidebar.
A new version downloads in the background and offers a one-click restart — no manual reinstall.
Attach text and image files to a task, up to 5 at a time — stays local unless your chosen provider is a cloud model.
Every embedded model, the custom-server option, and your own Claude API key — no feature is gated behind a paywall.
These are unsigned builds today, so your OS shows a one-time warning on first launch — normal for a beta, not a sign anything's wrong. On macOS: right-click the app → Open → confirm. On Windows: "More info" → "Run anyway." Only needed once.
A one-time payment that unlocks no feature — the app is already free and complete either way. It's just a way to say thanks if localagent has been useful to you.
Coming soonYes. Every embedded model, custom Hugging Face models, the custom-server option, and your own Claude API key are all included — nothing is gated behind a paywall. The optional one-time "Support the project" payment unlocks no feature; it's just a way to say thanks.
Not by default. The embedded and custom Hugging Face models run entirely in-process — zero backend, zero telemetry. Your code only leaves your machine if you explicitly choose Claude (sends task context to Anthropic's API with your own key) or point it at a custom OpenAI-compatible server.
macOS (Apple Silicon) and Windows (x64) today, as prebuilt installers. Builds are currently unsigned, so your OS shows a one-time warning on first launch.
No. The source is publicly viewable on GitHub, but it's proprietary — all rights reserved, no license is granted to copy, modify, or redistribute it. The prebuilt installers are free to use as distributed.
No. Signing in with Google is entirely optional and only used to back up session history to your own Drive — it gates no feature.
For support, please open an issue on GitHub.
Last reviewed: September 2026. Written to accurately describe what the app does — not a substitute for professional legal advice. Treat it as an honest draft, not a finished legal document.
There is no backend for this project — no server we operate, no analytics, no telemetry, nothing phoning home. Everything below happens directly between your machine and whichever service you've explicitly configured.
Your code and files stay on your machine unless you choose a provider that sends them elsewhere: Claude sends task context to Anthropic's API (needs your own API key); a custom server sends it to whatever OpenAI-compatible endpoint you point at. The default embedded mode — curated or a custom Hugging Face model — sends nothing anywhere once downloaded.
Session history is saved locally on your machine. It only leaves your machine if you sign in with Google, in which case it's backed up to a hidden, app-private folder in your own Google Drive — not visible in your normal Drive UI, not accessible to any other app or person.
Google sign-in is optional and gates nothing. If used, your email, name, and profile picture URL are requested from Google's own identity endpoint and stored locally; the refresh token is encrypted at rest via your OS's native credential store, never written to disk as plain text.
Crash and error logs are written locally and never uploaded anywhere — no external crash-reporting service, no account, no telemetry.
The optional "Support the project" purchase is processed entirely by the payment platform used to sell it — we do not receive or store your payment details ourselves. It unlocks no feature and creates no account.
Questions about this policy: lavuchandu169@gmail.com.
Last reviewed: September 2026. Written to accurately describe the terms this project intends — not a substitute for professional legal advice.
localagent is provided as-is, with no warranty of any kind, express or implied. You use it at your own risk, particularly in modes that let it edit files or run shell commands without per-action approval.
License. The source code is publicly viewable but proprietary — all rights reserved. No license to use, copy, modify, or redistribute the source is granted. The prebuilt installers are provided for personal use of the app as distributed.
The optional "Support the project" purchase does not create a contract for any feature, service, or ongoing obligation beyond the payment itself. Because it unlocks nothing and the app remains fully functional either way, all such purchases are final and non-refundable.
To the maximum extent permitted by law, the author is not liable for any damages arising from use of this software, including data loss or unintended file or system changes made by the agent while running with your permission.
Questions about these terms: lavuchandu169@gmail.com.