Skip to content
mattsaundPublic

About

A local AI multitool that fine tunes expert models, delegates to them, and allows circular computing

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

   ⠀⠀⠀⠀⠀⠀⢱⣆⠀⠀⠀⠀⠀⠀
   ⠀⠀⠀⠀⠀⠀⠈⣿⣷⡀⠀⠀⠀⠀
   ⠀⠀⠀⠀⠀⠀⢸⣿⣿⣷⣧⠀⠀⠀
   ⠀⠀⠀⠀⡀⢠⣿⡟⣿⣿⣿⡇⠀⠀
   ⠀⠀⠀⠀⣳⣼⣿⡏⢸⣿⣿⣿⢀⠀   C R U C I B L E
   ⠀⠀⠀⣰⣿⣿⡿⠁⢸⣿⣿⡟⣼⡆   An all in one local AI workshop.
   ⢰⢀⣾⣿⣿⠟⠀⠀⣾⢿⣿⣿⣿⣿
   ⢸⣿⣿⣿⡏⠀⠀⠀⠃⠸⣿⣿⣿⡿
   ⢳⣿⣿⣿⠀⠀⠀⠀⠀⠀⢹⣿⡿⡁
   ⠀⠹⣿⣿⡄⠀⠀⠀⠀⠀⢠⣿⡞⠁
   ⠀⠀⠈⠛⢿⣄⠀⠀⠀⣠⠞⠋⠀⠀
   ⠀⠀⠀⠀⠀⠀⠉⠀⠀⠀⠀⠀⠀⠀

Crucible builds software from a sentence, answers with the right expert for each question, and gets, makes and measures the models it runs. Local models do what they can, and frontier models do the heavy lifting once you add a key. Nothing leaves the machine until you say so.

Website: https://msaunders.dev/crucible/

One prompt, two experts: Qwen writes summary.html, then GPT OSS takes over for the README note


Install

From the releases page:

macOS Crucible-macOS.dmg drag into Applications
Windows Crucible-Setup.exe run it, no administrator needed
Linux Crucible-x86_64.AppImage chmod +x, run it

Not signed with a paid certificate. On Windows, click More info → Run anyway. On a Mac, open it once, then System Settings → Privacy & Security → Open Anyway.

The first start fetches the rest, with progress shown and nothing to click:

  • Crucible's own Python (no system Python needed or touched)
  • the compute runtimes the machine can use (CPU, CUDA, Metal, Vulkan)
  • the programs experts use, like GitHub's CLI, Node.js, a headless browser, Git on Windows and a text reader for pictures on Linux
  • with no model at all, an open coding model sized to the machine (Qwen, Apache-2.0), seated as Programming
  • the fine-tuner's environment

Or build from source

curl -fsSL https://raw.githubusercontent.com/mattsaund/Crucible/main/install.sh | bash
irm https://raw.githubusercontent.com/mattsaund/Crucible/main/install.ps1 | iex

The options are --prefix DIR, --jobs N, --check, --no-deps, --no-trainer, -y and --uninstall. Run from inside a clone, it builds that clone.

Updating: Settings → About → Update now, or run the installer again. Crucible checks GitHub once a day and says when there is one, and with Pre-releases ticked there it offers pre-releases and betas too. Config, models and history are kept. Versions are tagged v1.0.0 (--branch v1.0.0 builds one from source), and what changed is in CHANGELOG.md.

Uninstall: crucible --uninstall. It asks before deleting models. If the program will not run, use one of these.

curl -fsSL https://raw.githubusercontent.com/mattsaund/Crucible/main/install.sh | bash -s -- --uninstall
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/mattsaund/Crucible/main/install.ps1))) -Uninstall

What it does

Ask for anything

  • One box, one loop. Ask for anything. The expert works on it with the tools until it is done, and the delegator brings in another expert when the next step needs one.
  • A question is answered. Ask for software, like "a budgeting app that charts my gross and net revenue", and the expert reads, writes and runs what it takes, an action at a time, until the app is there.
  • The expert can change mid-prompt. When the next step wants another kind of expertise, the delegator picks who takes it, and that expert carries on with the same conversation.
  • Each action is shown above the reply, with the diff it made or the output it printed.
  • A change is a sentence. "Make the button blue" is made and shown.
  • Every prompt is a turn of the conversation, so "now add dark mode" knows the project.

Code

  • An editor beside the conversation, with tabs, a file tree, find and replace, search across the project, Ctrl+P to open a file by a few letters, and a terminal.
  • A file an expert changes while it is open reloads. Unsaved work is never written over.

The code editor, with the project's files beside it

Preview, tune, ship

  • Preview shows the page being made and reloads as an expert writes it.
  • Pick an element, then nudge padding, margin, fonts, colors and more. Apply to source hands the change to an expert, who writes it into the stylesheet.
  • Make it an app, a program of its own in Applications, the Start menu or the applications menu. It runs without Crucible and keeps its own data.
  • Source is git, with changes, diffs, commits, pull, push, history and CI runs. Publish to GitHub and Ship a release with gh.

The preview, with an element picked and its dashboard open

Experts

  • A roster of experts. A small router model, the delegator, sends each prompt to the expert that fits. It scores the roster, so it cannot name one that does not exist.
  • Pin a prompt to one expert from the box, or with /name.
  • Replies show Markdown, code and math, typeset from LaTeX.
  • Providers are Anthropic and anything that speaks the OpenAI API, like OpenAI, Gemini, DeepSeek, Kimi, OpenRouter, Groq or a server on your network.
  • The corner of the right-hand panel shows the models in memory, and the frontier model's tokens, cost at list prices, rate limits and context.

The roster: each seat with a description and a model

Tools

Every model, local or not, gets the same tools in a trusted folder:

  • LIST, READ, FIND, WRITE and EDIT files, PDFs, Word, sheets and pictures too
  • RUN commands in any shell, and PYTHON, GIT, GH, and START, LOGS and STOP for servers
  • SEARCH the web (DuckDuckGo, no key) and FETCH pages in a headless browser
  • RENDER an HTML page or SVG to PDF or PNG
  • TOOL calls any MCP server added in Settings
  • SCREENSHOT, CLICK, MOVE, TYPE, KEY and SCROLL on this computer, off by default
  • Edits wait for your OK unless Auto is on. RUN is not a sandbox.

The question Crucible asks once per folder

Attach

  • Attach documents, code, pictures and folders with +, Ctrl+U, or by dropping them on the window. Paste a picture with Ctrl+V.
  • PDFs, Office files, EPUB and text are read into the prompt. Pictures go to models that see, and the others get the words in them.

Models

  • Download from Hugging Face in Models. Search, sort by best match, downloads, likes or recently updated, pick GGUF or MLX and a quantization, and see whether it fits on the cards, runs partly on the CPU, or is too large. Downloads carry on after a dropped line or a restart and are checked against Hugging Face's checksum.
  • Each result shows its own picture when its card names one, and otherwise its maker's logo, the original maker's for a repack.
  • GGUF everywhere, MLX on Apple Silicon, where an MLX model can be the delegator too. LM Studio and Hugging Face folders work as they are.
  • Models that see pictures work when their mmproj projector sits beside them.
  • Models stay loaded while they fit, so a follow-up question loads nothing. When one has to make room, the one used longest ago goes, and one in use is never freed.

The Models tab: a repository's quantizations, each marked by how it fits

Create

  • Fine-tune an expert. Pick a base model and data, train it on your card (LoRA, or QLoRA with CUDA), test it and keep it. A tenth of the data is held out and scored before and after, so you see what training did.
  • Condense a model. A Q8 that does not fit becomes a Q4 that does, and the largest models go down to IQ2 or IQ1, made after a pass over sample text that finds the weights that matter most. Every size is measured first and marked by how it fits.

Condensing: every size, measured, and how each fits

The experts made here

Bench

  • Bench opens on the public board. The best local models come first, each with a Download that opens it in Models, then the spider chart, then every result there is, searchable.
  • Add benchmark runs your models through nine benchmarks in seven skills (MMLU-Pro, ARC-Challenge, GSM8K, MATH-500, HumanEval, MBPP, IFEval, HellaSwag and TruthfulQA). Quick asks 50 items of each, Standard 200, Full every one.
  • Every Crucible asks the same items the same way, so a score from one machine stands beside a score from another.
  • Spider charts compare up to three models. A result opens to every benchmark, its margin of error, and the answers that were wrong.
  • Post a result from your GitHub account. A result is names and numbers, nothing typed, and its model has to be a public file on Hugging Face.
  • Bench vs base puts a fine-tune beside the model it was made from.
  • HumanEval and MBPP run the code a model writes on this machine, with time and memory limits. It is not a sandbox.

Bench: the board's best models, a spider chart, and every result

Serve

  • Serve a model to other programs, the way LM Studio does. Turn it on in Settings, Server, pick the delegator or any model here, and point any OpenAI client at the address it shows.
  • It answers on this machine only, or on your network when you say so, over IPv4 and IPv6. An API key is optional.
  • Tools, pictures, JSON answers and embeddings work as OpenAI's clients expect. A provider's model can be served too, to other computers only with a key.
  • HTTPS when you turn it on, with your own certificate or one Crucible makes for this machine and keeps. Settings shows its SHA-256 to check it by. A certificate renewed in place is served as soon as it changes, and Crucible's own follows the machine to a new network.
  • Chats keep working while it serves. A request waits its turn.
  • It listens only while it is on.

The local server over HTTPS: its certificate's SHA-256, its addresses on this machine and the network, and the last requests

Hardware

  • Runtimes (CPU, CUDA, Metal, Vulkan) install from Settings, downloaded or compiled. A downloaded CUDA runtime needs only the NVIDIA driver, 580 or newer.
  • Several GPUs work as one, split automatically, evenly, by priority, or kept to a single card.

Runtimes: CPU, CUDA and Vulkan, all three active

History

  • History is per project, in the right-hand panel, and a reopened chat carries on where it stopped.
  • A chat with no project gets a folder of its own in ~/.crucible/Scratchpad (hidden). The right-hand panel lists recent chats and projects.

Every chat of a project


Settings and keys

Everything is in Settings and applies at once. The file is ~/.config/crucible/config.json, or %APPDATA%\crucible\config.json on Windows.

  • Ctrl+1 to 4 open Chat, Create, Models and Bench.
  • Ctrl+, opens Settings, and Ctrl+U attaches.
  • In the editor, Ctrl+P opens a file, Ctrl+Shift+F searches the project, Ctrl+S saves, Ctrl+F finds, Ctrl+H replaces and Ctrl+G goes to a line.

Build from source

C++20, CMake 3.24+, git, and on Linux WebKitGTK (libwebkit2gtk-4.1-dev on Debian and Ubuntu, webkit2gtk4.1-devel on Fedora). Dependencies are fetched and pinned. The tests want Python 3.10+ on PATH (or CRUCIBLE_PYTHON).

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)
sudo cmake --install build --component crucible

Contributing and AI policy

I am open to AI and agentic coding, but the code written needs to follow specific guidelines:

  1. MUST be human readable, acceptable variable/function names.
  2. Easily traceable, following a good program flow.
  3. Contributor MUST look at/document code and code changes. You need to understand the code that is being written.

Every push is built and tested on Linux, macOS and Windows. docs/testing says how to test a change or a release.


License

MIT, see LICENSE. What is compiled in is listed in THIRD_PARTY.md, and crucible --licenses prints the notices.

Some things are fetched rather than shipped, each under its own license. Crucible's Python (PSF License), the fine-tuner (mostly Apache-2.0, BSD and MIT, with NVIDIA's proprietary CUDA libraries on NVIDIA machines), the experts' programs, the starter model (Apache-2.0), the benchmarks' questions and every model's weights.

Provided "as is", without warranty. You are responsible for what your models read, write and run.

About

A local AI multitool that fine tunes expert models, delegates to them, and allows circular computing

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Contributors

Languages