JSPM

The knowledge-graph memory server for AI tools.
Mem0-grade intelligence. Engram-grade simplicity.

Quick Start · Architecture · Agent Setup · Contributing

What is ShelbyMCP?

Every AI memory server is a bag of embeddings. ShelbyMCP connects your thoughts.

ShelbyMCP is a zero-dependency MCP memory server with a built-in knowledge graph. It gives Claude Code, Cursor, Codex, Windsurf, Gemini, Antigravity, and any MCP-compatible AI tool persistent memory that understands how your thoughts are related — not just what they contain.

Ship it with the Forage skill, a scheduled task that runs on your existing AI subscription to continuously enrich, consolidate, and connect your memories. No Docker. No Python. No cloud accounts. Just a binary and a database file.

Why ShelbyMCP?

Quick Start

1. Install

# npx (no install needed)
npx shelbymcp

# Or install globally
npm install -g shelbymcp

# Or build from source
git clone https://github.com/Studio-Moser/shelbymcp.git
cd shelbymcp && npm install && npm run build

2. Set Up Your Agent

The CLI auto-configures everything — MCP server registration, Memory Protocol, and optional Forage skill:

shelbymcp setup claude-code --forage    # Claude Code CLI
shelbymcp setup claude-desktop --forage # Claude Desktop app
shelbymcp setup cursor --forage         # Cursor IDE
shelbymcp setup codex --forage          # OpenAI Codex
shelbymcp setup windsurf --forage       # Windsurf (Codeium)
shelbymcp setup gemini --forage         # Gemini CLI
shelbymcp setup antigravity --forage    # Antigravity (Google)

Drop --forage if you just want the MCP server without the scheduled enrichment skill.

That's it. The CLI registers the MCP server, adds the Memory Protocol to the right place, and installs the Forage skill. See docs/AGENT-SETUP.md for manual config and platform-specific details.

3. Add the Memory Protocol

Most agents get the Memory Protocol added automatically during setup. For agents that require manual steps, the CLI will tell you exactly what to do. Here's where it goes:

shelbymcp protocol >> ~/.claude/CLAUDE.md                              # Claude Code (auto)
shelbymcp protocol >> ~/.codex/AGENTS.md                               # Codex (auto)
shelbymcp protocol >> ~/.codeium/windsurf/memories/global_rules.md     # Windsurf (auto)
shelbymcp protocol >> ~/.gemini/GEMINI.md                              # Gemini CLI / Antigravity (auto)

For Cursor, paste into Settings > Rules > User Rules. For Claude Desktop, paste into Settings > General > "What personal preferences should Claude consider in responses?". The Memory Protocol tells your agent when to save and search — without it, the tools are available but won't be used proactively.

4. Seed Your Memory

Your database is empty after install. There are three ways to make it useful immediately:

# Option A: Run the onboarding interview (recommended)
# Paste this into a conversation — it asks a few questions and saves 15-30 memories
shelbymcp onboard

# Option B: Import from another AI tool
# Paste this prompt into ChatGPT/Claude/Gemini, copy the response back
shelbymcp migrate

# Option C: Just start working — memories accumulate naturally over time

The onboard skill runs a conversational interview covering who you are, what you're building, your team, preferences, and anti-patterns. Takes about 5 minutes. The migrate prompt tells your other AI tools to export everything they know about you in a structured format that ShelbyMCP can import.

5. Verify

Ask your agent: "What memory tools do you have available?"

It should list 9 tools. Then test: "Remember that I prefer dark mode in all my apps." — and in a new session: "What do you know about my preferences?"

How It Works

You (in Claude Code): "We decided to use CloudKit for sync instead of Firebase"

Claude Code → capture_thought tool → ShelbyMCP:
  1. Stores thought in SQLite
  2. Agent provides metadata (type: decision, topics: [sync, cloud])
  3. Agent suggests relationships to existing thoughts
  4. FTS5 indexes content for keyword search

Later:
You: "What did we decide about our sync strategy?"

Claude Code → search_thoughts tool → ShelbyMCP:
  1. FTS5 keyword search for "sync strategy"
  2. Returns thought + all connected thoughts via knowledge graph
  3. Agent has full context of decisions, alternatives considered, and related work

With the Forage Skill (scheduled, daily):

Forage runs on Claude Code's scheduler:
  1. Backfills embeddings for thoughts that don't have them
  2. Re-classifies poorly tagged thoughts
  3. Finds duplicate thoughts and merges them
  4. Detects contradictions ("we said PostgreSQL last month but SQLite this week")
  5. Discovers connections between thoughts across projects
  6. Sweeps for stale action items that fell through the cracks
  7. Generates a weekly digest of your thinking

MCP Tools

9 focused tools — research shows 5-8 tools per server is the sweet spot for agent accuracy.

Knowledge Graph

What makes ShelbyMCP different from every other memory server is the knowledge graph. Thoughts aren't isolated — they're connected.

Edge Types

Agents create edges at capture time ("this decision relates to thought X") and the Forage skill discovers additional connections over time.

The Forage Skill

ShelbyMCP ships with shelby-forage, a scheduled skill that runs on your existing AI subscription (Claude Code, Codex, etc.) to continuously improve your memory. The server stays zero-dependency — the intelligence comes from tools you're already paying for.

Setup

See docs/AGENT-SETUP.md for full setup instructions, platform compatibility table, and gotchas for each agent.

Without the Forage Skill

ShelbyMCP works fine without it — you get persistent storage, FTS5 search, and knowledge graph. The Forage skill adds the intelligence layer that makes memories smarter over time. Think of it as optional but recommended.

Getting Started: Onboard & Migrate

An empty memory database is a cold start problem — your AI tools can't personalize until they know something about you. ShelbyMCP ships with two tools to solve this:

The Onboard Skill (recommended)

A one-time conversational interview that seeds 15-30 foundational memories. Paste the prompt into a conversation with your primary AI tool:

shelbymcp onboard

It covers:

• Who you are — name, role, expertise
• What you're building — projects, goals, tech stack
• Who you work with — team, roles, stakeholders
• How you like to work — communication style, coding preferences, AI interaction style
• What to avoid — anti-patterns, pet peeves, past frustrations

Takes about 5 minutes. The skill adapts its questions based on your answers — if you're a solo founder, it won't ask about team structure. Memories are saved after each round so you can see them accumulate in real time.

The Migrate Prompt

Already have context stored in ChatGPT, Claude, Gemini, or another AI? Export it:

shelbymcp migrate

This prints a prompt you paste into your other AI tool. That tool dumps everything it knows about you in a structured format. Copy the response back into your ShelbyMCP-connected agent — the onboard skill will parse and import it, or you can just paste it into any conversation and ask the agent to import it.

Works with any AI that has memory or conversation history about you. Run it once per tool you're migrating from.

CLI Reference

shelbymcp                          Start the MCP server (stdio)
shelbymcp setup <agent>            Set up ShelbyMCP for an agent
shelbymcp setup <agent> --forage   ...and install the Forage skill
shelbymcp uninstall <agent>        Remove ShelbyMCP from an agent
shelbymcp protocol                 Print the Memory Protocol
shelbymcp forage                   Print the Forage skill prompt
shelbymcp onboard                  Print the onboarding interview prompt
shelbymcp migrate                  Print the migration prompt for other AI tools
shelbymcp help                     Show help
shelbymcp --version                Print version

Supported agents: claude-code, claude-desktop, cursor, codex, windsurf, gemini, antigravity

Server flags: --db <path> (custom database path), --verbose (debug logging)

See docs/AGENT-SETUP.md for manual config, platform-specific details, and setup for other MCP-compatible clients.

Architecture

ShelbyMCP is a single binary that communicates via MCP (stdio JSON-RPC) and stores everything in a single SQLite file.

AI Tool (Claude Code, Cursor, etc.)
    │
    │ MCP (stdio JSON-RPC)
    │
    ▼
┌──────────────────────┐
│      ShelbyMCP        │
│                        │
│  ┌──────────────────┐ │
│  │   MCP Protocol    │ │  ← JSON-RPC request/response
│  │   (stdio)         │ │
│  └────────┬─────────┘ │
│           │            │
│  ┌────────▼─────────┐ │
│  │   Tool Router    │ │  ← Routes to capture/search/link/etc.
│  └────────┬─────────┘ │
│           │            │
│  ┌────────▼─────────┐ │
│  │    SQLite DB      │ │
│  │  ┌─────────────┐ │ │
│  │  │  thoughts    │ │ │  ← Content, metadata, embeddings
│  │  │  thought_fts │ │ │  ← FTS5 full-text index
│  │  │  edges       │ │ │  ← Knowledge graph relationships
│  │  └─────────────┘ │ │
│  └──────────────────┘ │
└──────────────────────┘
         │
         ▼
    ~/.shelbymcp/memory.db  (single file)

See docs/ARCHITECTURE.md for the full technical design.

Comparison

Part of the Shelby Ecosystem

ShelbyMCP is the open-source memory server. Shelby for Mac (coming soon) is the native macOS app that adds:

• Always-on embedding pipeline (no scheduled skill needed)
• Instant auto-classification at capture time
• Semantic vector search
• CloudKit sync across all your Macs
• Heartbeat system (Pulse / Tidyup / Forage)
• Native menu bar + global hotkey quick capture
• Extension discovery and management

ShelbyMCP and Shelby for Mac use the same SQLite database. Start with the MCP server, upgrade to the Mac app when you want more.

Design Principles for Contributors

These are non-negotiable. They exist because MCP servers directly impact token costs for every user on every message.

1. Tool descriptions MUST be static. Tool definitions become part of the agent's system prompt and are sent on every message. Dynamic data (counts, timestamps, user-specific info) in descriptions breaks prompt caching — costing 10x more tokens. Put dynamic data in tool responses, not descriptions. See Architecture: Token Efficiency Patterns.

2. Search returns summaries, not full content. A search hitting 20 thoughts at 2,000 words each = 40K wasted tokens. Search results return the agent-provided summary field (one line). The agent calls get_thought for full content when it needs it.

3. All list/search tools have a limit parameter. Default 20, max 100. Responses include total_count and has_more. No unbounded queries.

4. The server runs zero inference. Agents provide metadata (type, topics, summary, relationships) at capture time. The Forage skill handles enrichment. The server is pure storage + retrieval.

5. All logging to stderr. console.error only. stdout is the MCP JSON-RPC channel. A single console.log breaks everything.

6. Keep the tool count focused. 9 tools. Research (Block, Phil Schmid, Docker) shows 5-8 per server is optimal. Consolidate related operations into single tools with action parameters.

7. Errors are instructions. Return isError: true with actionable messages. "Not found" is useless. "No thought with ID abc123. Try search_thoughts to find it by content." helps the agent self-correct.

8. Every tool gets annotations. MCP spec annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) on every registered tool. See Architecture: Tool Annotations.

Development

# Install dependencies
npm install

# Build
npm run build

# Run tests
npm test

# Run in development
npm run dev -- --db ./test.db --verbose

See docs/DEVELOPMENT.md for the full development guide.

Contributing

We welcome contributions! See CONTRIBUTING.md for the workflow.

• Issues first — Open an issue before starting work
• PRs welcome — Bug fixes, new MCP tools, documentation improvements
• Forage tasks — Propose new Forage skill tasks for memory enrichment

License

MIT - see LICENSE for details.

_{Built by Studio Moser. Your AI deserves a memory that connects the dots.}