A coding agent for your own GPU.
A real coding agent, driving a model on your machine. Local by default: your repo stays on the disk and nothing's metered, with the cloud there if you ever want it. One agent across the TUI, VS Code and desktop.
Plan, edit, run on a model you host.
01 · Get started
Download & point at your model
Grab LM Studio (or any OpenAI-style local server: llama.cpp, Ollama), load a model, flip on the local server, and paste the endpoint into Origami. Four steps, no keys, no account.
02 · Three surfaces
Just start talking to it
Open it where you work: VS Code is the priority, Desktop follows, the origami TUI is there today. Ask in plain English. It plans, edits, runs and shows the diff.
Under the hood
Workflow, built in
Run Origami out of a dedicated folder: desktop, Downloads, wherever suits. With your model connected, run /firstfold and it scaffolds the whole workspace for you: a projects folder, skills, crons and a wiki. A solid base to work from, and the structure that keeps a model's work organised instead of scattered.
/wrap is the end-of-session button. Like telling the agent to update its own handoff and wiki, it writes down what happened, so the model carries context from one session to the next and builds a working memory of your projects and its own actions. Pick up tomorrow where you left off.
Context matters for every model, and most of all on smaller windows. Origami tracks your turns and shows, right in the chat, how much of the window you've spent. At 90% it nudges you to run /wrap, preserving what the model has learned and keeping your handoff sharp before the window fills.
Everyone else builds for rented models. We don't.
A harness that is truly local-first, cloud-optional, because owning your compute is the only footing that lasts.
The problem
Everything funnels you to rented tokens
OpenCode is a great base. But, like almost every agent out there, it points you at hosted model-rental services. The whole category is tuned for a paid SaaS endpoint; the model already running on your own machine is an afterthought nobody optimises for.
Our stance
Local-first, cloud-optional, for real
We flip it. The GPU under your desk is the default path, tuned to be fast and first-class. The cloud is still one click away the moment you want it: opt-in, never assumed, never the thing we quietly nudge you toward.
Why it matters
Renting tokens doesn't scale
Today's token pricing is unsustainable, and as agents do more it only gets worse. Owning your own compute and hardware isn't a nice-to-have. It's imperative. This is the harness built for that world.
What that buys you
No meter, no lock-in
Building in the open.
A harness built to squeeze local models as far as they'll go: features kept clean and unconvoluted, built to run offline and compact.
Built on OpenCode (MIT) · full attribution ships with the release.
Documents unfolded for the new era.
The future of documents! Create in your browser, openable anywhere, completely local, no account, no subscription, FREE!
The deck is the document, and it opens locked.
01 · One file
Everything, in one .html
Slides, live blocks, real data, the styling: no server, no assets folder. Opens in any browser and runs offline, forever.
02 · Live blocks
Data, not pictures of it
A Gantt, a node graph, a sortable table, rendered in the page. Hand one fold to a model and it comes back snapped in, drift rejected.
03 · Security
A stranger's file opens locked
A deck you didn't make opens sandboxed and read-only. It shows its content, but privileged actions stay off until you unlock it.
- Renders & reads freelySlides, live blocks, charts, visible the moment it opens.
- Locked until you say soNetwork, storage and privileged actions stay off behind the padlock.
- You hold the keyOne deliberate unlock, per file. Never automatic, never for a stranger.
Be to the office what Blender is to video.
One substrate, three documents. We earn the next one, not assume it.
Seen a block we're missing? Send it.
Describe an idea, or paste HTML/SVG to become a future block. It composes a message you copy or email. This form sends and stores nothing on its own.
Tip
Paste beats describe
Hand us the HTML or SVG and it becomes a real block far faster. We lift it into the format and wire it up.
- You send itAn idea, or markup that already does the thing.
- We shape itInto a bounded, themeable Origami block.
- It shipsEveryone gets it next drop, offline, in one file.
Velkarn: The Darklands.
Game modding is a hobby of mine, and people think creatives will be out of a job. But guess what, executives aren't the ones that make games, creatives do. I'm out to prove it and create the means to show those who may be less sure on the tech how to produce their dreams.
A look at the dev kit
Sizzle reel · after dev-kit sign-offThe kit, at a glance
Four systems, running today
- 01Aether ForgeThe in-engine editor: terrain, weather, items, quests, and the mod kit that ships.
- 02AetherHumanHigh-quality faces, rigged and brought into the engine.
- 03AssetGenImage → 3D → in-engine, fast, filling a large world.
- 04Aether AIReactive enemy minds: perception + behaviour trees, real fights.
We're building the toolchain in the open.
The game will launch with a full, visually intuitive dev kit for modders and game-makers alike, in the Godot engine.
01 · Day-one dev kit
Aether Forge
The in-engine editor we build the world with is the mod kit that ships: terrain, placement, items, weather, quests, NPCs and AI, authored in Forge and handed to players on release.
02 · Content pipeline
AssetGen
Generate-to-game assets: image → 3D → in-engine, fast. How a small studio fills a large world, feeding straight into Forge.
From Tabletop to Screen.
The setting
Years in the making
Velkarn is a fantasy world, but not a new one. It's a setting I've built up in tabletop D&D for years, and this game will take place in a small part of it.
Pre-alpha, to followers first.
The near-term reveal goes to followers: the dev kit and tech first, the game as it earns the screenshots. No store page, no promises we can't show.