Small, sharp tools for working with language models. Some are single HTML files you can open from disk; one is a full app.
single-page-hosted-apps/turbine/ · full README
Feed Turbine a 1,000-page book and a prompt. It tiles the text into windows, sends each one to the model with read-only context on either side, validates the output, checkpoints every window, and stitches the results back into one document. Think of it as a pump: text goes in one end, transformed text comes out the other, with many repetitions in between.
- Bring your own inference. Ollama, vLLM and llama.cpp on your machine or LAN; OpenRouter (400+ models), OpenAI and Anthropic in the cloud. One connection page, one key vault, one countdown.
- Map or fold. Run windows in parallel, or sequentially with the previous output carried forward when continuity matters.
- Preview before you commit. Select any span of the source, even one line, and watch the model transform exactly that with the real prompt and context, streaming, with a live token counter.
- Tokens, not guesses. A Rust/WASM tokenizer in the browser, recalibrated by the model's own usage numbers after every preview, so the plan's budget and the backend's reality agree.
- Model parameters discovered, not assumed. Turbine asks each backend what the model accepts, shows those knobs, keeps reasoning off by default, and lets you turn it on where the model supports it.
- Conserve mode. A validator that strips Markdown and checks the words survived, so a formatting pass over an embeddings corpus keeps Darwin's prose and not the model's.
- Safe to host. Keys are encrypted in the browser under your passphrase and held server-side only in memory until the session timer runs out. A public deployment never fetches user-supplied URLs.
Built with DiamondJS on the front end and Elysia on Bun at the back. Building it shook out five DiamondJS bugs, all fixed in v2.2.3.
cd single-page-hosted-apps/turbine
bun install && bun run build && bun run start # http://localhost:7331
Self-contained HTML files in single-page-static-apps/. Open them in a browser; no server, no build.
| App | What it does | Try it |
|---|---|---|
| Tokens Per Second Simulator | Type a tokens-per-second rate, pick a corpus, and watch text stream at that speed. Useful for feeling what a model's throughput number means. | GitHub Pages |
| Text Stats Analyzer | Characters, lines and a whitespace token estimate for any pasted text. | GitHub Pages |
| OpenAI TTS Inspector | Poke at OpenAI's text-to-speech endpoint from a single page. | open openai_tts_inspector.html |
python-tools/— an Ollama CLI prompt utility and a tokens-per-second versus GPU-count optimizer.cli-tools/— shell helpers.single-page-hosted-apps/turbine/tools/fetch_corpus.ts— builds a 3×3 public-domain corpus (250 / 500 / 1,000-page tiers across neuroscience, an adjacent science and a literary control) from Project Gutenberg, normalized to Markdown without any LLM in the loop.
Issues and pull requests are welcome. Turbine's engine, server and client each have their own tests: bun test inside the turbine folder runs them all.
AGPL-3.0. See LICENSE.