⚠️ Security Warning: Crucible executes LLM-generated code via Code.eval_string with no sandbox. The model can access the full Elixir standard library, file system, network, and OS commands. Do not run Crucible with untrusted input in production environments without additional safeguards.

Crucible is a Recursive Language Model (RLM) execution engine for Elixir. It gives an LLM a stateful REPL — the model writes Elixir code, that code gets executed, the model sees the result, and the loop continues until the model arrives at a final answer.

How it works

1. You provide a question (what to solve) and input (data to work with)
2. Crucible sends both to an LLM provider
3. The model writes Elixir code, which Crucible evaluates in a stateful REPL
4. Results feed back to the model for the next iteration
5. The loop continues until the model sets final = "answer" or hits the iteration limit

Installation

As a library

Add to your mix.exs:

{:crucible, "~> 1.0"}

Standalone binary (Burrito)

Crucible uses Burrito for standalone binaries:

git clone https://github.com/Whoaa512/crucible.git
cd crucible
mix deps.get
MIX_ENV=prod mix release crucible

Binaries are output to burrito_out/ for your platform.

CLI Usage

crucible run "Summarize the key points" --input data.txt --provider openai --model gpt-4o
crucible skills list                    # View cached skill snippets
crucible skills clear                   # Clear skill cache
crucible logs list                      # List trajectory logs
crucible logs cleanup --max-age 7       # Delete logs older than 7 days
crucible providers                      # List available providers
crucible version                        # Print version
crucible --help                         # Show usage

Run options

Library Usage

{:ok, answer, meta} = Crucible.completion(
  "What is the average of these numbers?",
  "4, 8, 15, 16, 23, 42",
  provider: :openai,
  model: "gpt-4o",
  return_meta: true
)

Skill Caching

When --skills is enabled, Crucible stores successful (question, code) pairs in a SQLite database. On subsequent runs, it retrieves the top-3 most similar past solutions and injects them into the system prompt, helping the model solve similar problems faster.

Providers

License

MIT