MPM-Coding MCP

Turning AI Coding from "Demos" into "Delivery"

English | 中文

AI Coding That Survives Reality

AI coding is fun until you try it on a real repo:

The model forgets context ("where is the code?")
It edits based on guesses ("this should be fine")
Long tasks drift, skip steps, or die halfway
Later you cannot answer "what changed and why?"

MPM is not trying to make the LLM "smarter" or "better at chatting". That is the model's job. MPM makes the work finishable: locate first (code_search), check impact (code_impact), run long tasks as a phased chain with gates (task_chain), then store the why (memo).

In AI coding, "smart" usually means "steady": solves real problems, leaves a trail, can resume, with fewer guesses and fewer misses.

Even if your git history gets messy (or you rebuild the repo), the reasoning trail can still be there: memo stores the why in .mpm-data/. Back up .mpm-data/ and you can usually reconstruct faster (often cleaner) with AI.

What is MPM?

MPM is a set of MCP tools + rules for long-running, high-signal AI coding. Initialize once, then paste _MPM_PROJECT_RULES.md into your client's system rules.

🚀 30-Second Start (Do This First)

1) initialize_project
2) Paste _MPM_PROJECT_RULES.md into client system rules
3) Ask directly: "Help me fix XXX and follow the rules"

If you do this first, you can start effectively without learning every tool in advance.

Core Differentiators:

Traditional Approach	MPM Approach
`grep "some symbol"` → 500 results	`code_search("some symbol")` → exact file:line
"I think this change should work"	`code_impact` → full call chain analysis
Starting from scratch every session	`system_recall` → cross-session memory
Long tasks drift or stop halfway	`task_chain` → long-running task chain with gates

Practical Workflow: One Complete Loop (Example)

Below is a copy-paste ready example. Paste it into any MCP client to run.

Standard Mode (Recommended for Beginners)

Read _MPM_PROJECT_RULES.md and follow it.

Task: Fix <the issue you actually have>.
Requirements:
1. Locate the code first
2. Analyze impact scope
3. Implement the fix
4. Record the change reason

The AI will automatically execute: initialize_project → code_search → code_impact → modify code → memo to record.

Strict Mode (With Explicit Gates)

Read _MPM_PROJECT_RULES.md and follow it.

Use task_chain to complete the following task:
Task: Fix <the issue you actually have>.

Phase requirements:
1. Locate phase: Use code_search to find the target function
2. Analyze phase: Use code_impact to evaluate impact scope
3. Implement phase: Fix and pass tests
4. Wrap-up phase: Use memo to record change reason

Report results after each phase and wait for confirmation before proceeding.

Closed-Loop Checklist

Understand: project_map / flow_trace to grasp project structure
Locate: code_search to pinpoint symbols
Assess: code_impact to analyze call chain impact
Change: Write code, fix compilation errors
Verify: Run tests to confirm functionality
Record: memo to archive change rationale

⚠️ Data Hygiene: The .mpm-data/ directory stores local data and should not be committed to version control.

Project Binding: initialize_project creates .mpm-data/project_config.json as an anchor. Future sessions auto-bind to this project root. If multiple anchors are found under a workspace aggregator folder, MPM refuses to guess and requires explicit project_root.

What You Get

Find the right code faster (code_search, project_map, flow_trace)
Change with fewer surprises (code_impact)
Run long tasks with checkpoints (task_chain, system_hook)
Keep a usable change log (memo, system_recall)

Quick Start

1. Build

# Windows
powershell -ExecutionPolicy Bypass -File scripts\build-windows.ps1

# Linux/macOS
./scripts/build-unix.sh

2. Configure MCP

Point to the build output: mcp-server-go/bin/mpm-go(.exe)

3. Start Using

Initialize project
Help me locate and fix <your issue>, and follow _MPM_PROJECT_RULES.md

After initialization, MPM generates _MPM_PROJECT_RULES.md automatically. Treat it as the project's operating playbook:

It tells the LLM naming conventions, tool order, and hard constraints
You can start effectively without learning every tool detail first
In a new chat, ask the LLM to read this file first to reduce mistakes

Recommended first prompt: Read _MPM_PROJECT_RULES.md and follow it

4. Release Packaging (Fixed Directory)

python package_product.py

Notes:

Output directory is fixed: mpm-release/MyProjectManager
Each run removes previous mpm-release first, then rebuilds clean package contents

Documentation

MANUAL.md - Complete manual (all 13 tools + best practices + case studies)
MANUAL_ZH.md - 中文版完整手册
README_ZH.md - 中文版

Common Search Questions

How to do impact analysis in MCP? -> use code_impact
How to make LLM understand business logic flow? -> use flow_trace
How to monitor indexing progress for large repositories? -> use index_status
How to force full indexing? -> initialize_project(force_full_index=true)

See MANUAL.md for detailed examples.

OpenCode Multi-Agent Mode

MPM provides a 5-role Agent pack (PM / Architect / Coder / Expert / Spider) for direct use in OpenCode. See opencode-agents/README.md.

Contact

Support: GitHub Issues
Email: halflifezyf2680@gmail.com

License

MIT License

MPM-Coding