Mtrini Workspace

0.8 makes the agent work the way a coding agent should: it keeps its whole session, carries on after you approve a step, and now installs on Linux and keeps itself up to date.

• Updates itself

  New in 0.8.2. The app downloads new versions in the background, checks each against its published SHA-256, and installs it when you close the app. Or click Restart to update.

• Now on Linux

  New in 0.8.2: an AppImage that updates itself, plus a .deb and a .tar.gz. It opens in its own window with Chrome, Chromium, Edge, Brave or Vivaldi.

• A free model, built in

  Since 0.8.1 the agent works with no key, account or card: Qwen3.8 27B, sponsored by Heabsy, with a daily allowance that resets at midnight UTC.

• Plan mode and rewind

  The agent researches read-only, then shows a plan for you to approve. Rewind puts back the files it changed from any of your messages.

• Subagents, hooks and tasks

  Side jobs run in subagents with their own context. Hooks run your commands around tools. Longer jobs get a task list above the message box.

• Any number of custom providers

  Add as many OpenAI-compatible endpoints as you like, each with its own name, address and key. Only providers you have connected show in the model picker.

• A coding agent that checks its work

  A tool-using agent that reads your project, plans, edits, runs your real tests and verifies. It asks before it writes, runs commands, uses the network or does anything destructive.

• Local models

  Point it at llama.cpp, Ollama, LM Studio, vLLM or any OpenAI-compatible server. Local models stay on your machine.

• 20+ API providers

  OpenAI, Anthropic, Gemini, OpenRouter, Mistral, DeepSeek, Groq and more. Switch models from the top bar without restarting.

• Hugging Face built in

  Search, inspect and download models, then import them into a local model registry.

• Editor, terminal and Git

  Monaco editor, an integrated terminal, Git status and diffs, and a command menu on Ctrl+K.

• MCP tools

  Connect any MCP server over stdio or streamable HTTP, inspect its schemas, and lend its tools to the agent.

• Project memory and skills

  Per-project instructions and skills live in .mtrini/ alongside your code, with global skills shared across every project.

• No telemetry

  API requests go only to providers you configure. Credentials live in the OS keychain, never in config, git, logs or frontend code.

Every request runs through the same loop, shown live in the session view. Type / in the composer for the built-in commands (/help, /clear, /skills, /mcp and /settings) and for your own skills.

1. 01Understand→
2. 02Plan→
3. 03Inspect→
4. 04Use tools→
5. 05Modify→
6. 06Test→
7. 07Fix→
8. 08Verify→
9. 09Respond

# 1. Download MtriniWorkspace-Setup-0.8.2.exe (link above)
# 2. Run it, or install silently:
MtriniWorkspace-Setup-0.8.2.exe /SILENT

# 3. Open Mtrini Workspace, pick a folder,
#    add a provider in Settings, start a session

# Serve a Mtrini GGUF locally
llama-server -m Mtrini-SVL-1.1-BF16.gguf --port 8000

# Then in Settings → Providers → local
#   Base URL  http://localhost:8000/v1
#   Model     the id your server reports

git clone https://github.com/compiwerai/mtrini-workspace.git
cd mtrini-workspace
pip install -r backend/requirements.txt
npm install --prefix frontend

# backend (http://localhost:8787)
python -m uvicorn mtrini.main:app --app-dir backend --reload --port 8787

# frontend (http://localhost:5173)
npm run dev --prefix frontend

To build the Windows app and installer yourself, run scripts/build-exe.ps1, then iscc scripts/setup.iss with Inno Setup 6.