Minder

Minder is a self-hosted MCP (Model Context Protocol) platform for repository-aware engineering intelligence.

It runs natively on macOS, Linux, and Windows — no Docker, no external services. The stack is a Python FastAPI server with SQLite, Turbovec (embedded vector search), and llama-cpp-python for local LLM inference, distributed as a Tauri desktop app or standalone server.

What's in this repo

Architecture

Developer workstation
  ├── minder-cli          repo sync, MCP config
  └── AI agent (IDE)  ──► Minder Server :8800
                              │
                  ┌───────────┼───────────────────┐
                  │           │                   │
               SQLite    Turbovec           llama-cpp-python
          (users, sessions,  (embedded vector    (LLM, GGUF,
           workflows, graph)  search, .tvim file)  Metal/CPU)

• Transport: SSE (/sse), streamable HTTP (/mcp), stdio
• LLM inference: llama-cpp-python with GGUF models auto-downloaded from HuggingFace (Metal on Mac, CPU elsewhere)
• Vector search: Turbovec — embedded, file-based 4-bit quantized ANN index, no separate server
• Relational storage: SQLite (default) or PostgreSQL
• Desktop: optional Tauri v2 shell that bundles the Python server as a sidecar

Quick Start

1. Install and run

To install the official desktop app release directly:

• macOS / Linux:

  curl -fsSL https://github.com/hiimtrung/minder/releases/latest/download/install-minder-release.sh | bash
  

• Windows 10/11:

  iwr -useb https://github.com/hiimtrung/minder/releases/latest/download/install-minder-release.ps1 | iex
  

For local source development:

# Install dependencies and build the dashboard
make native-install

# Start the server (dashboard + MCP API on port 8800)
make native-run

GGUF models (ggml-org/gemma-4-E2B-it-GGUF) are downloaded automatically on first startup. No manual setup required.

2. Open the dashboard

http://localhost:8800/dashboard

First run → /dashboard/setup to create an admin and get your mk_... key.

3. Create a client and connect your IDE

1. Open /dashboard/clients → create a client → save the mkc_... key (shown once)
2. Open the client detail → copy the MCP snippet for your IDE from Copy-ready MCP snippets
3. Open /dashboard/instruction → copy the agent orchestration rules for your IDE

4. Install the CLI and sync a repository

# Install
uv tool install minder-cli

# Log in
minder login --client-key mkc_your_key --server-url http://localhost:8800/sse

# Write MCP config to your IDE (optional — dashboard shows it too)
minder install --target vscode --target claude-code

# Sync a repository
minder sync

MCP Tools (26 tools)

Minder exposes a lean set of 26 tools organized into 7 groups. All session tools are always available; other tools require explicit tool_scopes on the client.

Auth

Session (always available)

minder_session_boot is the single entry point for every session flow — it creates, finds, or restores a session in one call.

Memory

Skills

Workflow

Search & Graph

Agents

Session Continuity

minder_session_boot is the single session entry point for all agents. It handles create, find, and restore transparently:

# First run on a project — creates a new session
minder_session_boot(project_name="my-api", project_context={"repo_path": "/dev/my-api"})
→ { session_id: "a1b2...", session_found: false }

# After /compact or machine switch — recovers by name
minder_session_boot(project_name="my-api")
→ { session_id: "a1b2...", session_found: true, session_summary: {...}, _next_steps: [...] }

# Checkpoint work and update branch context in one call
minder_session_save(
  session_id="a1b2...",
  state={"task": "...", "next_steps": [...]},
  branch="feat/my-feature",
  open_files=["src/service.py"]
)

Dashboard Pages

Documentation

• Local Dev Setup
• Production Deployment
• System Design
• Admin & Client Onboarding
• Minder CLI
• MCP Tool Reference

License

Apache License 2.0. See LICENSE for details.