MCP 服务器

iGenius Memory MCP Server

Provides persistent AI memory and visual analysis tools for any MCP-compatible agent via the iGenius Memory service.

README

iGenius Memory — Persistent AI Memory for Any Agent

A structured, AI-powered memory backend that gives any MCP-compatible agent persistent memory via the iGenius Memory service. All AI processing happens server-side — you just need an API key.

3 Ways to Use iGenius

Client	Install	Best For
🧩 VS Code Extension	Marketplace	Full sidebar UI, memory browser, AI provider settings
⚡ MCP Server	`pip install igenius-mcp`	Any MCP client — VS Code, Claude Desktop, Cursor, Windsurf
🖥️ Desktop App	Windows Installer	Standalone system-tray app, works with any editor

Get a free API key at igenius-memory.online — all three clients use the same key.

1. VS Code Extension (Marketplace)

Install directly from the VS Code Marketplace — no pip, no config files:

ext install igenius-memory.igenius-memory

Or search "iGenius Memory" in the Extensions panel. Includes sidebar UI, memory browser, status bar indicator, AI provider settings, and auto-warms briefings on a configurable interval.

2. MCP Server (pip)

For any MCP-compatible client (VS Code Copilot, Claude Desktop, Cursor, Windsurf, etc.):

pip install igenius-mcp

Then add to your MCP config:

VS Code — ~/.vscode/mcp.json:

{
  "servers": {
    "igenius-memory": {
      "command": "igenius-mcp",
      "env": { "IGENIUS_API_KEY": "ig_your_key_here" },
      "type": "stdio"
    }
  }
}

Claude Desktop / Cursor / Windsurf — add to your MCP config file:

{
  "mcpServers": {
    "igenius-memory": {
      "command": "python",
      "args": ["-m", "igenius_mcp.server"],
      "env": { "IGENIUS_API_KEY": "ig_your_key_here" }
    }
  }
}

⚠️ Windows users: If VS Code can't find igenius-mcp, use python -m igenius_mcp.server instead.

3. Desktop App (Windows)

Standalone system-tray application — works alongside any editor or IDE:

Download Installer (NSIS setup or MSI)
Built with Tauri + Rust — lightweight, native, ~5 MB
System tray with quick access to briefings, search, and memory stats
Configure LLM provider (LM Studio, OpenAI, Anthropic, Google) from the UI

Restart VS Code after installing the extension or adding MCP config — all 17 memory tools become available to Copilot and any MCP-compatible agent.

Available Tools

Tool	Description
`memory_briefing`	Session briefing from all memory layers (call FIRST)
`memory_ingest`	Ingest user/agent messages for AI extraction
`memory_consolidate`	Merge accumulated extracts into master briefing
`memory_process`	Detect trigger words and auto-classify text
`memory_store`	Direct store to a specific memory layer
`memory_search`	Natural language search across memories
`memory_recall`	Retrieve all persistent session memories
`memory_summarize`	LLM-powered summary of a memory layer
`memory_delete`	Delete a memory by ID
`memory_update`	Update fields on an existing memory
`memory_review`	List short-term memories for triage
`memory_promote`	Promote short-term → long-term
`memory_pin`	Pin a fact permanently (user-confirmed, never expires)
`memory_triggers_list`	List trigger words and their layers
`memory_triggers_add`	Add a new trigger word
`visual_report`	Render URL → screenshot → vision analysis → full UI/UX report (requires `[visual]`)
`visual_screenshot`	Render URL → return base64 PNG (requires `[visual]`)

LLM Requirements

iGenius uses an LLM backend for AI extraction, consolidation, and (optionally) visual analysis. You can use a local or remote LLM provider.

Local Setup (LM Studio, Ollama, etc.)

Requirement	Minimum
GPU VRAM	6 GB+
Recommended model	Qwen 3.5 4B (non-thinking) or equivalent
Context window	3,000+ tokens

⚠️ IMPORTANT: Do NOT use thinking/reasoning models (e.g. QwQ, DeepSeek R1, o1, o3). Thinking models emit <think> chains before the actual response, which breaks iGenius's structured JSON extraction pipeline. Only use standard non-thinking (instruct/chat) models.

Why these specs? iGenius sends structured extraction prompts that expect clean JSON output. A 4B-parameter non-thinking model at 3k context is the sweet spot for fast, accurate extraction without hallucination or timeouts. Larger models (8B+) work too — just ensure you have the VRAM headroom and that the model is a non-thinking variant.

Remote Setup (OpenAI, Anthropic, Google, etc.)

No local hardware requirements. Any API-accessible model works — configure the provider, model name, and API key in the VS Code extension settings or environment variables.

Environment Variables

Variable	Required	Default
`IGENIUS_API_KEY`	Yes	—
`IGENIUS_API_URL`	No	`https://igenius-memory.online/v1`

Visual Tools (Optional)

Give your AI agent eyes — render any URL, take a pixel-perfect screenshot, and get instant UI/UX analysis from a local vision model.

Install

pip install "igenius-mcp[visual]"
python -m playwright install chromium

Then load a vision-capable model in LM Studio (e.g. Qwen 3.5 9B Vision, non-thinking).

⚠️ Do NOT use thinking/reasoning vision models — same restriction as above.

Visual MCP Tools

Tool	Description
`visual_report`	Render URL → screenshot → vision analysis → full UI/UX report
`visual_screenshot`	Render URL → return base64-encoded PNG (no analysis)

Visual Environment Variables

Variable	Default	Description
`IGENIUS_VISION_URL`	`http://localhost:1234/v1`	Vision model API endpoint
`IGENIUS_VISION_MODEL`	auto-detect	Override the vision model name
`IGENIUS_VISION_KEY`	—	API key for vision endpoint (e.g. LM Studio auth token)
`IGENIUS_VIEWPORT_W`	`1280`	Screenshot viewport width
`IGENIUS_VIEWPORT_H`	`800`	Screenshot viewport height

100% local — screenshots and analysis never leave your machine.

Agent Instructions

For best results, add the iGenius agent instructions to your workspace:

VS Code: Place igenius.instructions.md in ~/.vscode/prompts/
Claude Code: Add to CLAUDE.md
Workspace: Add to .github/copilot-instructions.md

Get the template at igenius-memory.info

How It Works

Agent ←→ MCP (stdio) ←→ igenius-mcp ←→ REST API ←→ iGenius Backend

The memory tools are a thin proxy — they translate MCP tool calls into REST API requests. All AI extraction, LLM summarization, and encryption happens server-side.

The visual tools run locally — Playwright renders URLs on your machine and a local vision model (e.g. LM Studio + Qwen2.5-VL) analyzes the screenshots. Screenshots and analysis never leave your machine.

Plans

Plan	Price	Requests	API Keys	IPs/Key
Starter	Free	1,000/week	1	3
Pro	$19/mo	50,000/day	5	10
Enterprise	Contact	500,000/day	20	50

Details at igenius-memory.store

Coming Soon

iGenius Context Engine — unlimited effective context for local LLMs through intelligent recursive summarization. Run a 3B model with a 4K context window and handle conversations of any length.