arete-mcp-server

OpenIdentity-compatible MCP server for Claude Desktop. Makes Claude remember who you are — across conversations, across tools.

What is OpenIdentity?

OpenIdentity is a portable identity protocol for AI. One schema that works across Claude, GPT, Cursor, and any AI tool.

Arete is the reference implementation.

Quick Start

# 1. Sign up (just need your email)
npx arete-mcp-server setup

# 2. Configure Claude Desktop (~/.config/claude/claude_desktop_config.json):
{
  "mcpServers": {
    "arete": {
      "command": "npx",
      "args": ["arete-mcp-server"]
    }
  }
}

# 3. Restart Claude Desktop

# 4. Ask Claude: "What do you know about me?"

That's it. No repo cloning, no building.

Commands

npx arete-mcp-server setup              # Interactive signup + install Claude Code hooks
npx arete-mcp-server setup EMAIL        # Non-interactive signup
npx arete-mcp-server                    # Start MCP server
npx arete-mcp-server extract PATH       # Extract facts from transcript (used by hooks)
npx arete-mcp-server --help             # Show help

Tools Provided

| Tool | Purpose | |------|---------| | arete_identity | Get your identity for personalization (supports task-aware projection) | | arete_remember | Store, validate, or remove facts (auto-detects category) | | arete_activity | Get recent browsing/interaction context | | arete_infer | Extract facts from activity patterns + accept/reject candidates | | arete_onboard | Interactive interview to build identity (3-5 min) |

How It Works

Your identity is stored locally (~/.arete/) in OpenIdentity format. The MCP server exposes it to Claude Desktop, so Claude knows:

• Who you are (role, background)
• What you're working on (current projects, focus areas)
• Your preferences (communication style, tools you use)
• Your expertise (skills, domains)

Export & Import

Your identity is portable. Backup, transfer, or share:

# Export your identity
arete identity export --format oi > my-identity.oi

# Export only public facts (for sharing)
arete identity export --format oi --visibility public > public.oi

# Import from backup or another tool
arete identity import ./my-identity.oi

Privacy Tiers

Facts have visibility levels you control:

| Tier | Description | |------|-------------| | public | Safe for any AI tool | | trusted | Only authorized apps (default) | | local | Never leaves device |

Recommended: System Prompt

For Claude to automatically capture context (not just respond to manual tool calls), add this to your Claude Desktop system prompt:

Settings → Claude's memory & system prompt → paste this:

You have Arete tools for portable identity.

AT CONVERSATION START: Call arete_identity to know who you're talking to.

AFTER RESPONSES: If the user revealed something DURABLE — useful in future conversations — store it with arete_remember:

Personal examples:
- "I'm Brazilian" → arete_remember(content: "User is Brazilian")
- "I prefer short answers" → arete_remember(content: "Prefers concise responses")
- "I'm learning Rust" → arete_remember(content: "Currently learning Rust")

Project/strategic examples (equally important!):
- Strategic pivot discovered → arete_remember(content: "Project shifting to X approach")
- Architecture decision → arete_remember(content: "Uses Y pattern for Z reason")

Category is auto-detected (core, expertise, preference, context, focus).

Test: Would a future AI benefit from knowing this? If yes, store it.

Without this, you get tools but Claude won't proactively use them.

Claude Code: Automatic Context Capture

For Claude Code users, setup auto-installs hooks that make context capture automatic — no system prompt needed:

npx arete-mcp-server setup
# ✅ Detects Claude Code and installs hooks to ~/.claude/settings.json

Hooks installed:

| Hook | When | What | |------|------|------| | SessionStart | New conversation | Injects "call arete_identity" instruction | | PreCompact | Before context compression | Extracts facts via Haiku | | SessionEnd | Session ends | Final extraction of durable facts |

How it works:

1. Hooks read the conversation transcript (JSONL)
2. Call Claude Haiku to extract durable facts (role, skills, preferences)
3. Merge into ~/.arete/identity.json with semantic deduplication

Check extraction logs: cat ~/.arete/extraction.log

FAQ

How is this different from ChatGPT/Claude memory?

Those memories are siloed. OpenIdentity is the protocol that connects them — same you, everywhere. One identity that works across Claude, GPT, Cursor, and whatever comes next.

Privacy concerns?

Local-first by design. Your identity lives in ~/.arete/ on your machine. Code is open source. You control visibility tiers.

Why MCP?

MCP (Model Context Protocol) is Anthropic's standard for extending Claude. As MCP adoption grows, OpenIdentity automatically works with any MCP-compatible tool.

Specification

Full protocol spec: OPENIDENTITY.md

Open Beta

No invite code needed — just run npx arete-mcp-server setup and enter your email.

License

MIT