You don't edit anymore. Adits does.
adits.ai · live, in public beta · try the hosted version
Adits is a workspace for making things with an AI coding agent. You drop in files, describe what you want, and the agent does the actual work — writing HTML, editing images, cropping PDFs, rearranging slides, whatever the project needs. A new file lands back on disk. You iterate.
If you've tried Claude Design, the surface will feel familiar: a chat on the left, a canvas on the right. Adits takes the same idea and rebuilds it around two principles.
Claude Design ships an assistant that lives inside Anthropic's product. Adits doesn't ship one. Every turn hands off to a coding CLI running on the same machine: Claude Code, Gemini CLI, Codex, or OpenCode. The agent owns the filesystem, the edits, the tool calls, the multi-turn reasoning.
Adits is the prompt IDE. Every surface — the drop zone, the inline tweak knobs, the element picker, the draw overlay, the questions form, the skill chips — is a way to author the next instruction for the agent without having to type it out. The agent is the runtime; Adits is how you compose the turn.
That seam buys you two things:
- Bring your own CLI, your own subscription. If
claudeis already logged in on your machine, Adits uses it. No second bill, no captive model, no proxy in the middle. - Swap the agent per turn. Prefer Gemini for one kind of edit and Claude for another? Pick the executor in the composer. The rest of the app doesn't care which one runs.
Adits ships in two modes. The hosted mode runs on our infra. The self-hosted mode runs entirely on your laptop — one Node process, one local directory, no cloud.
In self-hosted mode there is no Clerk account, no provisioned VM, no webhook round-trip, no Redis, no managed database. Every project is a plain folder under ~/.adits/projects/<id>/. Your edits happen in place, against the files sitting on your disk. Sharing a project is tar czf. Backing one up is cp.
git clone <repo-url>
cd adits
pnpm install && pnpm build
# Either `claude` is already logged in, or:
export ANTHROPIC_API_KEY=sk-ant-...
export ADITS_BACKEND=local
node --import tsx server/index.ts
# open http://localhost:4001That is the whole install. No Postgres to run, no Redis, no Docker, no account to create.
Claude Design is scoped to HTML-shaped artifacts. Adits opens anything you can drop into a folder, and every file type ships its own viewer and its own in-tab toolbar:
| Type | What you get |
|---|---|
| HTML pages | Landing pages, wireframes, slide decks, dashboards, posters. Full editing lanes — Tweaks, Edit, Draw, Comment. |
| Images | Canvas viewer with View / Marker / Draw / Crop. Crop exports a PNG you can hand back to the agent. |
| PDFs | Page-by-page viewer with draw overlays; crop a region to PNG. |
| Office docs | .docx, .xlsx, .pptx render in an embedded editor. |
| Audio / video | Native player plus a timeline strip; segment a range to reference in the next turn. |
| Text / Markdown / CSV | Inline render, no external dependency. |
Sketches (.napkin) |
Pen, shapes, arrows, sticky notes — hand the agent a rough drawing instead of typing. |
Adding a new file type is one component — register it, and every tab, chip, and thumbnail in the app picks it up automatically.
Two nouns, that's it:
- Project — a folder on disk. One project holds one body of related work.
- Page / file — anything inside it. The unit of work.
Two panes, that's it:
- Chat (left) — the conversation with the agent. It is the project's history: every turn that produced the files you're looking at.
- Bench (right) — the project's files, rendered as tabs. A permanent
Design Filestab lists everything; clicking a file opens it in its own tab.
One rule: the chat holds the conversation, the bench holds the files. A file never appears in the chat. A chat turn never appears in the bench. Every gesture you make on the bench side (adjusting a knob, clicking a button, drawing a stroke, leaving a comment) enqueues a chip into the composer — you stack chips, add typed context, and send the whole batch as a single reviewable turn.
You can always just type. You usually won't need to:
- Tweaks — inline knobs the agent wires up on demand. Pull a slider, flip a toggle, pick a color. Values stay in memory until you hit Save, then they get packaged into one turn.
- Edit — click an element on the page and change typography / sizing / layout in a properties panel. Each change applies live and queues a structured instruction for the agent.
- Draw — a full-screen canvas overlay. Scribble strokes, drop sticky notes, circle the bit you want changed. On Send, the overlay goes along with the file.
- Comment — pin a free-form note to an element. Keep it as feedback for yourself or "Send to Adits" to promote it into an agent instruction.
- Skills — pre-made recipes the agent can invoke. "Make a deck," "Wireframe," "Interactive prototype," etc. Pick one from the composer and the agent loads the matching playbook for this turn.
- Open the Welcome Workspace. A starter project is seeded on first login with a handful of sample files — a deck, an image, a PDF, a Markdown note. You have something to click on immediately.
- Drop a file into the chat or the bench. Uploads land in the project's
uploads/folder. The agent can read any of them on the next turn. - Ask for a deck.
"Make me a 5-slide deck about <topic>."The agent picks up themake-a-deckskill, writes an HTML file that works standalone in any browser, and drops it intoDesign Files. - Tweak it. Click into the deck's properties. Pull a slider. Circle a title with Draw. Add a sticky note. Hit Send. The agent applies everything in one turn.
- Export it. Page → HTML bundle. Deck → PPTX or PDF. Image → PNG. Your files are your files.
That's the whole shape. Project in, files out, an agent you control in the middle.