Inference-first Unsloth Studio fork: Chat + MCP + Settings.

[!NOTE] This repository variant is configured as an inference-first Studio fork. Training and export APIs/routes/CLI commands are removed from the active product surface. Supported focus areas are chat/inference, model loading, MCP integration, multimodal chat input, and TTS output.

What This Version Is • UI Overview • Features • Quickstart • MCP Secrets

🎯 What This Version Is

This fork is a streamlined Studio build oriented around local and hybrid inference workflows.

• Top-level product tabs are Chat, MCP's, and Settings.
• Chat supports GGUF and non-GGUF inference backends.
• Local model discovery is configurable from Settings.
• MCP providers are managed via persisted JSON config and wired into chat tool-calling.
• Training and export flows are intentionally out of scope for this variant.

🖼️ UI Overview

Chat

MCP

Settings

⭐ Features

Inference + Chat

• Run local and remote models from a single chat interface.
• GGUF + non-GGUF model loading support.
• Multimodal input support for text, images, and file attachments.
• Optional voice dictation input and TTS output path.
• Chat sidebar with folders, drag-drop organization, and compare mode support.

MCP

• MCP provider configuration via ~/.unsloth/studio/mcp.json.
• Provider + tool enable/disable controls from the chat tools panel.
• MCP tools participate in chat tool-calling flow when enabled.
• MCP config test workflow from the MCP page.

Settings

• Configure default local model source path.
• Rescan and inspect discovered local models.
• Local model candidates are prioritized in model selection UI.

⚡ Quickstart

From this repository (recommended)

./scripts/studio.sh setup
./scripts/studio.sh run -H 0.0.0.0 -p 8000

PowerShell:

.\scripts\studio.ps1 -Command setup
.\scripts\studio.ps1 -Command run -Host 0.0.0.0 -Port 8000

Via installed CLI

pip install --upgrade pip uv
uv pip install -e . --torch-backend=auto
unsloth studio setup
unsloth studio -H 0.0.0.0 -p 8000

🔐 MCP Secrets and Safety

• mcp.json can contain sensitive values if you inline keys. Avoid that.
• Prefer env references in MCP config values:
  • $MY_SECRET
  • ${MY_SECRET}
• For HTTP MCP providers, prefer api_key_env over inline api_key.
• This repo ignores MCP config secret files via .gitignore:
  • **/mcp.json
  • **/mcp.local.json
  • **/mcp.secrets.json

🔗 Links

• Unsloth Documentation
• Unsloth Studio Docs
• Unsloth Discord
• Unsloth Model Catalog

Citation

@software{unsloth,
  author = {Daniel Han, Michael Han and Unsloth team},
  title = {Unsloth},
  url = {https://github.com/unslothai/unsloth},
  year = {2023}
}

License

Unsloth uses a dual-licensing model of Apache 2.0 and AGPL-3.0.

• Core package: Apache 2.0
• Studio UI and related components: AGPL-3.0