No3sis - Knowledge Engine MCP Server

MCP server exposing the Synapse Pattern Map to Claude Code agents

No3sis (νόησις - "understanding, knowledge") is a FastMCP-based server that wraps the Synapse System's knowledge engine, enabling AI agents to access collective intelligence stored in Neo4j, Redis, and vector embeddings via the Model Context Protocol (MCP).

Architecture

Claude Code Agents
    ↓ (MCP stdio protocol)
No3sis MCP Server (FastMCP)
    ↓ (subprocess wrapper)
Synapse Knowledge Engine CLI Tools
    ├─ synapse_search.py → Neo4j + BGE-M3 vectors
    ├─ synapse_standard.py → Pattern Map standards
    ├─ synapse_template.py → Project templates
    └─ synapse_health.py → Infrastructure monitoring
        ↓
Infrastructure
    ├─ Neo4j (Pattern Map storage - 247+ patterns)
    ├─ Redis (Corpus Callosum cache)
    └─ BGE-M3 (Semantic embeddings)

The 4 Knowledge Tools

1. search_pattern_map

Search the Synapse Pattern Map for relevant patterns, solutions, and best practices.

Claude Code Usage:

mcp__no3sis_search_pattern_map(
    query="error handling patterns rust",
    max_results=5
)

2. get_coding_standard

Retrieve language-specific coding standards from the Pattern Map.

Claude Code Usage:

mcp__no3sis_get_coding_standard(
    standard_type="naming-conventions",
    language="rust"
)

3. get_project_template

Access project templates and boilerplate code.

Claude Code Usage:

mcp__no3sis_get_project_template(
    template_type="cli-app",
    language="rust",
    variables='{"project_name": "my-app"}'  # Optional JSON string
)

4. check_system_health

Check health of Synapse knowledge engine infrastructure.

Claude Code Usage:

mcp__no3sis_check_system_health()

Installation

Prerequisites

• Python 3.12+
• Running Synapse knowledge engine (Neo4j + Redis)
• Synapse tools installed at ~/.synapse-system/.synapse/neo4j/
• uv or pip for installation

Setup

# Navigate to no3sis directory
cd /path/to/no3sis

# Install with uv (recommended)
uv venv
uv pip install -e .

# OR install with pip
pip install -e .

# Configure environment
cp .env.example .env
# Edit .env to set SYNAPSE_NEO4J_DIR

Environment Configuration

Create/edit .env file:

# Path to Synapse knowledge engine tools
SYNAPSE_NEO4J_DIR=/home/username/.synapse-system/.synapse/neo4j

# Optional: Default max results for searches
MAX_RESULTS_DEFAULT=10

# Optional: Enable debug logging
DEBUG=false

Registering with Claude Code

Method 1: Using claude mcp add (RECOMMENDED)

This is the correct way to register the MCP server with Claude Code CLI:

# Register in local scope (project-specific)
claude mcp add no3sis \
  "/path/to/no3sis/.venv/bin/python" \
  --scope local \
  --transport stdio \
  -e "SYNAPSE_NEO4J_DIR=/home/username/.synapse-system/.synapse/neo4j" \
  -- -m no3sis.server

# Verify registration
claude mcp list
claude mcp get no3sis

Scope options:

• local - Only available in current project (stored in ~/.claude.json for project)
• user - Available globally for all projects
• project - Uses .mcp.json in project root (requires approval)

Method 2: Project .mcp.json (Alternative)

Create .mcp.json in your project root:

{
  "mcpServers": {
    "no3sis": {
      "command": "/path/to/no3sis/.venv/bin/python",
      "args": ["-m", "no3sis.server"],
      "env": {
        "SYNAPSE_NEO4J_DIR": "/home/username/.synapse-system/.synapse/neo4j"
      }
    }
  }
}

Note: Project-scoped .mcp.json files require explicit user approval for security.

Verification

# Check if server is connected
claude mcp list
# Should show: no3sis: ✓ Connected

# Test in interactive session
claude
# Type: /mcp
# Should show no3sis with 4 tools available

Testing

CLI Testing Mode

The server supports direct CLI invocation for testing (bypasses MCP protocol):

cd /path/to/no3sis

# Test health check
.venv/bin/python -m no3sis.server health

# Test pattern search
.venv/bin/python -m no3sis.server search "error handling" 5

# Test coding standards
.venv/bin/python -m no3sis.server standard naming-conventions rust

# Test templates
.venv/bin/python -m no3sis.server template cli-app rust

MCP Protocol Testing

# Check MCP server health
claude mcp list

# Get detailed server info
claude mcp get no3sis

# Test in Claude session
claude
# Then use the tools via MCP protocol

Integration with Claude Code Agents

Agent Definition

Update .claude/agents/boss.md (or other agents):

---
tools: Read, Grep, Glob, Write, Bash, mcp__no3sis_search_pattern_map, mcp__no3sis_get_coding_standard, mcp__no3sis_get_project_template, mcp__no3sis_check_system_health
---

## Instructions

Use `mcp__no3sis_search_pattern_map` to query the Pattern Map for relevant patterns.
Use `mcp__no3sis_get_coding_standard` to retrieve language-specific standards.
Use `mcp__no3sis_get_project_template` to access project templates.
Use `mcp__no3sis_check_system_health` to check infrastructure status.

Tool Naming Convention

When Claude Code registers MCP tools, they're prefixed with mcp__<server_name>_:

• search_pattern_map → mcp__no3sis_search_pattern_map
• get_coding_standard → mcp__no3sis_get_coding_standard
• get_project_template → mcp__no3sis_get_project_template
• check_system_health → mcp__no3sis_check_system_health

Architecture Details

MCP Protocol Implementation

No3sis uses FastMCP with stdio transport:

• FastMCP handles MCP protocol (JSON-RPC over stdio)
• Tools are registered as async functions with @mcp.tool() decorator
• Communication happens via stdin/stdout pipes
• Claude Code spawns the server process and communicates bidirectionally

Subprocess Wrapper Strategy

The server shells out to existing Synapse CLI tools:

• Advantage: Zero changes to Synapse codebase
• Advantage: Easy to maintain and debug
• Overhead: ~20ms per request (acceptable for knowledge queries)
• Alternative: Future optimization could import Synapse modules directly

Request Flow

1. Agent invokes: mcp__no3sis_search_pattern_map("rust error handling", 5)
2. Claude Code → MCP JSON-RPC request via stdio → No3sis FastMCP
3. No3sis → subprocess.run(synapse_search.py) → Synapse tool
4. Synapse tool → Neo4j/Redis query → return JSON
5. No3sis → parse JSON → return to FastMCP
6. FastMCP → MCP response via stdio → Claude Code
7. Agent receives results

Why FastMCP?

• Simplicity: Decorator-based tool registration
• Protocol Compliance: Full MCP protocol support
• Async: Non-blocking I/O for multiple requests
• Stdio Transport: Standard for Claude Code integration
• Type Safety: Python type hints for tool parameters

Why Separate Repo?

• Independent versioning: No3sis can evolve separately from Synapse
• Reusability: Other projects can use the same knowledge engine
• Clean boundaries: MCP protocol is infrastructure, not agent logic
• Publishable: Can become pip install no3sis-mcp-server

Directory Structure

no3sis/
├── README.md              # This file
├── LICENSE                # MIT
├── pyproject.toml         # Python package config (dependencies: mcp, python-dotenv)
├── .env.example           # Config template
├── .gitignore
├── src/no3sis/
│   ├── __init__.py       # Package exports
│   └── server.py         # MCP server (~270 lines)
│       ├── FastMCP setup
│       ├── 4 tool implementations (@mcp.tool decorators)
│       ├── CLI testing interface
│       └── stdio transport (run_stdio_async)
└── tests/
    └── test_integration.py

Troubleshooting

Server Not Connecting

Error: no3sis: ✗ Failed to connect

Solutions:

1. Check venv path: ls /path/to/no3sis/.venv/bin/python
2. Test CLI mode: .venv/bin/python -m no3sis.server health
3. Check environment: Ensure SYNAPSE_NEO4J_DIR is set correctly
4. Re-register: claude mcp remove no3sis -s local && claude mcp add ...

Tool Not Found Errors

Error: Synapse tool not found: synapse_search.py

Solutions:

1. Verify SYNAPSE_NEO4J_DIR in .env
2. Check tools exist: ls $SYNAPSE_NEO4J_DIR/synapse_*.py
3. Ensure Synapse is installed and configured

Connection Errors (Neo4j/Redis)

Error: Neo4j service not accessible

Solutions:

# Check services are running
docker ps | grep -E "neo4j|redis"

# Restart services
cd ~/.synapse-system/.synapse/neo4j
docker-compose up -d

Tools Not Available in Claude

Issue: /mcp shows "No MCP servers configured"

Solutions:

1. Ensure you're in the correct project directory
2. Check registration: claude mcp list
3. Restart Claude Code session
4. Verify .claude.json contains no3sis entry

Development

Running Tests

# Install dev dependencies
uv pip install -e ".[dev]"

# Run integration tests
pytest tests/ -v

Adding New Tools

1. Add tool function in server.py:

@mcp.tool()
async def my_new_tool(param: str) -> str:
    """Tool description for Claude."""
    result = _run_synapse_tool("synapse_newtool.py", [param])
    return json.dumps(result, indent=2)

1. Create corresponding Synapse CLI tool
2. Update agent definitions with mcp__no3sis_my_new_tool

Performance

• Latency: ~20-50ms per request (subprocess overhead)
• Throughput: Async, handles multiple concurrent requests
• Caching: Synapse tools implement Redis caching (300s TTL)
• Future: Direct Python imports could reduce latency to ~5ms

Contributing

This repo is designed to be portable - it can live in the Synapse workspace or as a standalone repo at github.com/no3sis-lattice/synapse.

Pull requests welcome for:

• Performance optimizations (direct imports vs subprocess)
• Additional Synapse tools
• Improved error handling
• Documentation improvements

License

MIT