Generative UI for the coding agent
you already trust.

Yes — a browser interface for the terminal coding agent you already use, rendered faithfully: same engine, same settings, memory, and permission rules as your terminal. Install with npm i -g mirafold, run mirafold in a project directory, and your agent opens in the browser with generative UI on top — live cards, tables, charts, and dashboards instead of a wall of scrollback.

Yes. A session is fully local when the agent behind it points at a local inference server — Claude Code against Ollama’s Anthropic-style endpoint, or Codex against Ollama, LM Studio, or vLLM. Then nothing — code, key, or tokens — leaves your machine. The repo documents the exact recipe and an honest model/hardware table.

In your local daemon’s environment, exactly as in a terminal — ANTHROPIC_API_KEY and friends. It is never serialized to the browser, and on Pro the relay carries end-to-end-encrypted frames plus the connection metadata any forwarder handles (IPs, timing, byte counts) — never the key or the conversation.

Claude and Gemini: no — their terms prohibit driving a subscription login from a third-party app, so Mirafold refuses those and takes an API key instead, enforced in code, not fine print. A Codex/ChatGPT sign-in works locally, with the honest caveat shown at sign-in: OpenAI hasn’t clearly permitted third-party use in writing, but visibly tolerates it — your account, your call. No subscription of any kind runs over the relay. Local and open models have no restriction at all.

Two zones, hard boundary. The agent’s output renders only into an output zone: streamed markdown (no raw HTML), schema-validated registry components, and — for arbitrary UI — a sandboxed iframe with an opaque origin and a CSP that cuts every network path. The prompt box, socket, and credentials live in a trusted shell the agent cannot render into, and anything a component does back is validated server-side against an allowlist.

Claude Code, Codex, and Gemini CLI ship today, each behind its own adapter — none privileged. A new agent is one adapter behind a small interface, not a rewrite; the adapter spec in the repo documents exactly what an engine needs to qualify.

Yes — run as many sessions as you want, each with its own directory and agent. Mission control shows every session in one grid, at a glance, and you can act on any of them straight from the grid — no hunting through terminal tabs or tmux panes.

They work, in the same page. Type ! before any command and it runs in a real PTY on your machine, so interactive prompts behave: sudo’s password prompt appears in the shell’s own masked input — the agent never sees what you type — and afterward the agent reads the command’s output and picks up from there.

Yes — that’s Pro: pair a phone, tablet, or any browser with a QR code, and the same generative-UI session travels with you, pinned widgets and all. The relay in between is end-to-end encrypted — it sees ciphertext and bare connection metadata, never your key, your code, or your content. The relay’s code is open source (MIT) like everything else; the hosted relay is the convenience Pro pays for.

The whole product is MIT — shell, daemon, all three adapters, and the relay code itself, so you can read exactly what the piece you’d be trusting can and cannot see. Pro is convenience infrastructure: the hosted relay that lets your devices reach your sessions from anywhere, end-to-end encrypted. If you’d rather self-host the relay, you run the same code we do.