fused-render

Download for macOS, Windows, and Linux →

A local file explorer for your whole computer. Browse any directory in the browser, preview files, and author your own interactive views: any .html file you open gets a tiny injected runtime that can call a Python main() function and sync its state to the URL.

Runs entirely on 127.0.0.1. No accounts, no cloud, no sandboxing — your own machine, your own trusted code. See SPEC.md / ARCHITECTURE.md / DECISIONS.md for the full design.

Right-click a file in Explorer → Open with → fused-render, and it opens in your browser. See Windows: Explorer "Open with" to enable it.

Install

macOS app — the packaged FusedRender.app (bundles the fused CLI and rclone; no Python required):

brew install --cask fusedio/tap/fused-render

or download the DMG from the releases page.

Python package — each release also attaches a wheel (see the release notes for its URL): pip install <wheel-url>. From a source checkout:

pip install -e .

Requires Python 3.10+. Installs FastAPI, uvicorn, and pyarrow (used by the built-in parquet preview).

Building from source and the local dev loop live in CONTRIBUTING.md.

Run

fused-render

Opens a browser tab at http://127.0.0.1:1777/, starting in your home directory. Useful flags:

fused-render --start-dir ~/data --port 9000 --no-browser

--start-dir only sets the initial location — the whole filesystem stays browsable from there.

Windows: Explorer "Open with"

fused-render-open --register

Registers fused-render into Explorer's right-click "Open with" menu (HKCU only, no admin) for every format it previews — double-clicking a file, or picking "fused-render" from Open With, reuses a running server or starts one detached, then opens the file. fused-render-open --unregister removes the associations.

Authoring model

Any .py file is a runnable target as long as it defines a main() function:

# sine.py
import math

def main(n: int = 80, freq: float = 1.0):
    return {"points": [[i / n, math.sin(2 * math.pi * freq * i / n)] for i in range(n)]}

Any .html file can call it and bind the result to the URL:

<input id="freq" type="range" min="0.1" max="5" step="0.1" />
<script>
  const slider = document.getElementById("freq");
  slider.addEventListener("input", () => fused.params.set("freq", slider.value));
  fused.params.onChange(draw);

  async function draw() {
    const freq = fused.params.get("freq") || "1.0";
    const { points } = await fused.runPython("./sine.py", { freq });
    // ...render points...
  }
  draw();
</script>

• fused.runPython(pyPath, params) — runs main(**params) of a local .py file in a fresh subprocess and returns its JSON result. pyPath may be relative (to the HTML file) or absolute.

• fused.params — a string-only key/value store synced into the browser URL (get, getAll, set, onChange). Refreshing or bookmarking a view reproduces its exact state.

• fused.ai(prompt, opts?) — ask an AI model through the claude (Claude Code) CLI on your machine; resolves with {text, model, usage}. Pass a Hugging Face repo id as model ("Qwen/Qwen3.5-4B") and the same call runs a model locally instead — the slash is what tells them apart. Local chat works on every supported desktop platform: MLX is preferred on Apple Silicon, and everywhere else the default is a small CPU-only PyTorch build that runs on any machine — CUDA and ROCm builds exist and are one radio away on the AI Models page's Engines tab, offered only where the app can see a matching GPU. The AI Models page suggests models that suit whichever backend is serving you. Local calls also take history (prior {role, content} turns, for a conversation rather than one question) and can be stopped mid-answer with fused.ai.cancel().

• fused.ai.image({prompt, ...}) — text to image, locally; resolves with the PNG's path, a ready-made URL to point an <img> at, and the seed used (one is chosen for you if you don't pass one, so a render is always repeatable). It runs for minutes, so onProgress fires per denoising step — each tick with a previewUrl for a thumbnail of the image so far, so there is a picture emerging rather than a number going up — and the download manager's ✕ really stops it.

• fused.ai.transcribe({path, ...}) — speech to text, locally: point it at an audio or video file on this machine and it resolves with the words plus the {start, end, text} segments that carry their timestamps. task picks between transcribing in the original language and translating to English; the language is auto-detected unless you name one. It runs for minutes, so onProgress fires with seconds of audio and the download manager's ✕ really stops it, and the transcript is written to a file so it outlives the tab.

• fused.ai.models.list() / load(id) / unload(id) — what this machine is holding in memory and what it costs. See the AI Models page (docs).

• fused.trackJob(spec) — report work your page is doing that runs longer than the page itself (exporting a few thousand tiles, converting a folder) to the download manager in the bottom-right corner, so it stays visible after you browse away:

  const job = fused.trackJob({ title: "Export tiles", kind: "export",
                          unit: "items", cancellable: true });
  job.update({ done: 1200, total: 4096, detail: "zoom 12" });
  if (job.cancelRequested) stopTheWork();   // the manager's ✕ asked
  job.finish("Exported");                   // or .fail(err) / .cancelled()

  Reporting never throws and never blocks: a failed report cannot break the work it describes. Your page is the only thing that can stop its own work, so the ✕ here is a request you honour by checking cancelRequested.

• fused.watchJob(id) — the other half: watch work the server is doing (fused.ai.image(), fused.ai.transcribe() and fused.ai.models.load() all hand you an id) with .onUpdate(cb), .get() and a .cancel() that really stops it — the server owns those processes, so its ✕ is an act rather than a request.

Built-in preview templates (parquet tables, images, text/code files) are themselves just HTML files built on these same two primitives — open fused_render/templates/ to see how.

Configuration & advanced usage

Runtime features and settings live in docs/usage.md:

• Execution engine — built-in subprocess runner vs. the fused local compute backend (folder-level pyproject.toml dependencies in cached venvs), and FUSED_RENDER_ENGINE.
• Remote storage (mounts) — mount S3-compatible stores, Google Drive, and anything else rclone speaks, as local folders.
• AI Models — what the Hugging Face cache holds on this machine, what it costs on disk, how to clear it, and a search of the Hub that says which results you already have, beside a short list of suggested models you can download.
• Preferences — the in-app settings panel (execution engine, logs, template registry).
• Export for hosted serving — the programmatic POST /api/export bundle format for publishing a page yourself.

Claude Code plugin

This repo doubles as a Claude Code plugin marketplace. Installing the plugin adds skills that teach Claude how to use a fused-render project (running the explorer, opening views by URL), author fused-render views (the fused.runPython bridge, URL-synced params, file IO helpers), and build custom preview templates.

From inside Claude Code:

/plugin marketplace add fusedio/fused-render
/plugin install fused-render@fused-render

Or from the command line:

claude plugin marketplace add fusedio/fused-render
claude plugin install fused-render@fused-render

The manifests live in .claude-plugin/ (marketplace.json + plugin.json); the skills themselves are under skills/.

Installing is optional and only affects your own Claude Code sessions. Chats started from inside fused-render (the Claude and split-view templates, and app scaffolding) already get these skills: the app assembles the same plugin under ~/.fused-render/skill-plugin/ and loads it per session with --plugin-dir, so they work on a plain wheel or DMG install with nothing cloned.