ChaosNexus

Appearance

About ChaosNexus Crucible ​

ChaosNexus Crucible is the local LLM interface for the ChaosNexus stack. It loads a model through a pluggable inference backend and exposes a small HTTP API for generation. It can also open an MCP client toward ChaosNexus Anvil so a generation request may discover Anvil tools.

Interface, not weights

Crucible talks to models. It does not ship or train model weights. Training and datasets live in ChaosNexus Tuned. The default download is the published GGUF TunedChaos/ChaosNexus_Tuned_v1-GGUF, cached under ~/.chaosnexus/crucible/models. The PEFT adapter remains at TunedChaos/ChaosNexus_Tuned_v1.

Privacy (Hugging Face) ​

When you set a Hugging Face token in Forge Settings → Models, Forge stores it locally in app settings.toml and injects it as HF_TOKEN into the Crucible process only. Pull/download requests go to Hugging Face Hub. Tuned Chaos does not receive that token or your model traffic.

What it is ​

  • A Rust service (Axum) that binds locally (default 127.0.0.1:8080).
  • Backends selected by config: Candle (default, GGUF) or Colibri.
  • Optional MCP client connection to an Anvil instance when /generate includes Anvil port + token.
  • Model HTTP API: /models/status, /models/pull, /models/list.

What it is not ​

  • Not a desktop chat UI (Forge Agent Chat + Models settings own that).
  • Not the fine-tuning pipeline (Tuned).
  • Not a substitute for Anvil's sandbox; Anvil remains the tool host.

Status ​

Alpha. Default path is Tuned v1 GGUF via Candle. See the roadmap M6 checklist.

Where it lives ​

Next ​