IDE-Multi-Agent-Protocol

English | 中文说明

Zero-Dependency, Fully Local, Context-Persistent Multi-Agent Architecture for IDEs.

Are you tired of your IDE AI Agent losing context during long coding sessions? Do you hate passing API keys to heavy frameworks like CrewAI or Mem0 just to simulate multi-agent collaboration?

IMAP (IDE Multi-Agent Protocol) is a lightweight, zero-API architecture designed explicitly for native IDE Agents (like Antigravity / Cursor / Copilot). It uses a "Token Handoff" mechanism back by a local MCP SQLite database to achieve immortal memory and robust persona switching.

---

1. Zero External APIs: No OpenAI keys, no Anthropic keys. It purely leverages your IDE's built-in conversational models for reasoning, using Python's standard sqlite3 and the open-source mcp protocol for state management.
2. Context Continuity (Immortality): Solves the "context length limit" problem. When your chat gets too long, trigger /handoff. The agent summarizes its state, saves it to the SQLite brain, and gives you a [Token]. Paste the token into a fresh chat window, and it literally "wakes up" exactly where it left off, retaining 100% of its persona and project context.
3. True Separation of Concerns: Unbloats the "Global Prompt". We split the AI into three physical layers:

• Layer 3 (Router): Global Manager Rule analyzing intent.
• Layer 2 (Agents): Project-agnostic SKILL folders containing distinct personas (e.g., Writer, Engineer).
• Layer 1 (Memory): A persistent background MCP server tracking global decisions.

---

graph TD;
    User["👶 User"] -- "/dispatch" --> Manager["🧠 Layer 3: Manager / Router (Global Rule)"]
    Manager -- Generates --> Token["🎫 Hand-off Token"]
    Token -. User pastes into new chat .-> NewChat["💬 New IDE Chat Window"]
    NewChat -- "/switch" --> Persona["🎭 Layer 2: Expert Persona loaded fro