gnn-mcp-protocol - SKILL.md Agent Skill

name: gnn-mcp-protocol description: GNN Model Context Protocol processing and tool registration. Use when registering GNN operations as MCP tools, building MCP server configurations, or integrating GNN capabilities with LLM tool-use workflows.

GNN MCP Protocol (Step 21)

Purpose

Processes Model Context Protocol (MCP) configurations and registers GNN pipeline operations as MCP tools. Enables LLM agents to invoke GNN capabilities through standardized tool interfaces.

Key Commands

# Run MCP processing
python src/21_mcp.py --target-dir input/gnn_files --output-dir output --verbose

# As part of pipeline
python src/main.py --only-steps 21 --verbose

API

from mcp import (
    MCP, MCPRegistry, MCPServer, JSONRPCServer, MCPTool, MCPResource,
    create_mcp_server, start_mcp_server,
    register_tools, register_module_tools,
    get_available_tools, list_available_tools,
    list_available_resources, process_mcp,
    get_mcp_instance, handle_mcp_request,
    initialize,
)

# Process MCP step (used by pipeline)
process_mcp(
    target_dir, output_dir,
    verbose=True,
    performance_mode="high",
    strict_validation=True,
    cache_ttl=120.0,
)

# Create and start a JSON-RPC server bound to the global MCP registry
server = create_mcp_server()      # returns JSONRPCServer
start_mcp_server()

# Discover and register all pipeline modules' tools on the global MCP
register_tools()                             # populates mcp_instance
register_module_tools("gnn")                 # named module only

# Query available tools and resources
tools     = get_available_tools()            # same as list_available_tools()
tools_md  = list_available_tools()
resources = list_available_resources()

Key Exports

MCP — core registry (tools, resources, modules, performance metrics).
MCPRegistry — alias of MCP, the core registry.
MCPServer — JSON-RPC 2.0 request handler bound to an MCP instance.
JSONRPCServer — JSON-RPC 2.0 request handler bound to an MCP instance.
MCPTool / MCPResource — dataclasses backing the registry.
create_mcp_server / start_mcp_server — construct and start a JSON-RPC server.
register_tools — discover and register all pipeline modules' tools on the global MCP.
register_module_tools — register a single named module.
initialize — check SDK + discover modules; exposes performance_mode, enable_caching, enable_rate_limiting, strict_validation, cache_ttl, modules_allowlist, per_module_timeout, overall_timeout, force_refresh.
process_mcp — pipeline entry point (forwards overrides to initialize).
get_available_tools, list_available_tools, list_available_resources.
Error types: MCPToolNotFoundError, MCPValidationError, MCPModuleLoadError, MCPInvalidParamsError, MCPToolExecutionError, MCPSDKNotFoundError, MCPResourceNotFoundError, MCPPerformanceError.

Configuration

The MCP singleton honours these knobs, all propagated through initialize() and process_mcp(**kwargs):

Knob	Type	Default	Purpose
`performance_mode`	`"low"`/`"high"`	`"low"`	Bulk toggle for caching, rate limiting, strict validation
`enable_caching`	`bool`	from mode	Enable result cache
`enable_rate_limiting`	`bool`	from mode	Enable per-tool rate limiting
`strict_validation`	`bool`	from mode	Enforce JSON-schema validation of params
`cache_ttl`	`float` (sec)	`300.0`	Result-cache TTL
`modules_allowlist`	`list[str]`	`None`	Only load these modules under `src/`
`per_module_timeout`	`float`	`30.0`	Max seconds per module during discovery
`overall_timeout`	`float`	`120.0`	Wall-clock budget for parallel discovery
`force_refresh`	`bool`	`False`	Re-discover modules even if already loaded

Output

MCP configuration in output/21_mcp_output/
Tool registration manifests
Server configuration files

MCP Tools

This module registers tools with the GNN MCP server (see mcp.py):

None

References

AGENTS.md — Module documentation
README.md — Usage guide
SPEC.md — Module specification
MCP Protocol Spec

Documentation

README: Module Overview
AGENTS: Agentic Workflows
SPEC: Architectural Specification
SKILL: Capability API