runs on your computer · asks before it acts · nothing hidden
LoClaw is a Windows desktop app that runs AI on your own computer, using your own models. Your code, your drawings, your documents — nothing gets uploaded, because there's nowhere to upload it to.
It can chat, review code, propose edits, run commands, and drive the real apps on your desktop. It shows you what it's about to do, then waits for you to say yes.
No cloud. No account. No telemetry.
for files you'd rather keep to yourself
Source code. Customer drawings. Contracts. Financial records. A client's project you signed an agreement about. Work you'd simply rather not hand to a company you've never met.
Right now people with those files have two options: skip AI entirely and fall behind, or use it quietly and hope nobody asks. LoClaw is the third option. The model runs on your machine. The files stay on your machine. Nothing is uploaded, because there's nowhere to upload it to.
No cloud AI tool survives that. LoClaw keeps working — chat, agent mode, skills, connections to your local apps — because none of it was ever going anywhere. You don't have to take our word for it. That's the point: it's a thing you can check yourself, on your own hardware, in about ten seconds.
Ask for help with a client's drawings, a private repo, an invoice — without handing the file to somebody else's chat service. It never leaves the folder it's in.
No account. No cloud workspace. No hidden telemetry. Nothing quietly syncing in the background because it decided your files would be useful somewhere.
Sealed Mode refuses every path off the machine — not warned about, refused. For anyone who wants it that way, and for firms that need to lock it on so it can't be switched off.
And to be straight with you: no software can make your company compliant — that's a conversation for you and whoever signed the agreement. What LoClaw gives you is the thing cloud AI can't, which is a workflow where the default path for your files isn't the internet.
control is the feature
LoClaw doesn't blur the line between talking and acting. Chat only is the default, and it can't touch anything. Agent mode is a deliberate flip of a switch — and even then, every action asks first.
Ask questions, review ideas, draft code, explain errors. No file reads, no edits, no commands, no tool calls. There's nothing to approve because LoClaw isn't doing anything to approve.
Read files, propose edits, run commands, use approved connections. You see the diff before a write, the command before it runs, and the output after.
Ask a different local model to review an answer before you trust it. Peer review, without sending your code anywhere.
offline skill packs
Ask a local model about a niche workflow and it will often sound confident while missing the details that decide whether the thing actually works. CAD APIs, ESP32 wiring, DXF units, PowerShell, bookkeeping — the specifics generic models fumble.
A skill is a few kilobytes of plain markdown that tells the model the truth: the real API, the real units, the real mistakes. Skills ship with the app but stay off. You read one first — it's markdown, open it in Notepad — then you decide. Nothing reads what you typed and phones somewhere to pick a skill for you.
If it can't be emailed, copied to a USB stick, and read in Notepad, it doesn't belong in LoClaw.
local tools, visible connections
LoClaw can connect to the real apps on your computer — CAD, spreadsheets, editors — through an open standard called MCP. The point isn't that it can call them; plenty of things can. The point is that you see the app, the destination, the exact command, and the arguments before anything runs.
hop into autodesk fusion's MCP and generate me 3 random spheres
I'll generate 3 random spheres in the active Fusion model using TemporaryBRepManager.
A real run. A local model read the live design, wrote correct Fusion API Python, asked permission, executed it inside the running app, and checked its own work. Nothing was uploaded — there was nothing to upload to. That last line isn't our slogan; it's what the app printed when it finished. The same pattern works for any desktop app that speaks MCP.
the trust model
Anything can claim privacy. These are specific, checkable refusals — because a promise you can't inspect isn't a promise.
Point it at a model server on your network and the status line changes to stayed on your local network. Point it somewhere remote and it says that instead. That line is a fact, not a slogan.
"Allow for session" is scoped to the exact file, the exact command, the exact tool and arguments. Approving one write doesn't quietly bless the next one.
A project you clone can ship a config asking LoClaw to launch tools or switch model servers. It won't. You're shown the exact command line, and you decide.
Every front-end library is bundled in the app, and the chat view runs under a policy that blocks outbound requests. The air-gap is enforced, not asserted.
No confidence scores, no made-up percentages. And when a model can't actually call tools, LoClaw stops you rather than letting it pretend it tried.
where it actually is
It builds, it runs, and it did the thing above. It's also young, and this page isn't going to pretend otherwise.
One email when there's something worth installing. No drip campaign, no newsletter, no tracking pixel.