Skip to content

MCP Server Projection

coreason-meta-engineering is air-gapped from coreason-runtime. Downstream agents CANNOT import its Python functions directly. Instead, all AST manipulation logic is securely projected over the Model Context Protocol (MCP).

Launching the Server

The server is implemented using fastmcp and can be initialized seamlessly via the CLI. To expose the capability set over stdio for agent interaction, simply run:

uv run coreason-meta-mcp

Available Tool Projections

The MCP server exposes three strict tools for generating different classifications of assets. Each asset explicitly requires an action_space_id conforming to the URN constraints.

1. scaffold_manifest_state

Used to deterministically inject passive state models into the AST map. It requires: - state_name: The class name of the Pydantic state model. - geometric_schema: The raw JSON defining the fields. - target_file_path: Path to the module to mutate. - action_space_id: Uniquely identifiable cryptographic URN.

2. scaffold_logic_actuator

Used to instantiate an execution function within the runtime domain. - actuator_name: The runtime function to forge. - geometric_schema: The JSON schema defining function parameters. - target_file_path: The module to perform AST node insertion in. - action_space_id: Execution URN boundary mapping.

3. scaffold_epistemic_node

Used to engineer new Swarm Agent entities constrained to specific cognitive boundaries. - node_name: Name of the Agent Class. - cognitive_boundary_directive: Text defining the bounding epistemic policy. - target_file_path: Output file. - action_space_id: Routing URN to intercept payloads.