AgenticSystemCore

What AgenticSystemCore MCP server does

The AgenticSystemCore MCP server connects an MCP client to a local folder of Markdown files called a Bundle. Each file contains typed fields, and the reference engine turns the folder into a knowledge node with readable content, graph data, search-related files, agent-facing text, and discovery information.

The server is intended for agents that need to work from a controlled knowledge base. It can search and read items, follow links between them, produce citations by address, compose selected content into a starting harness, and retain information through the Bundle's proposal-based governance model. The material is local: the README describes the MCP mode as operating without sending content away from the machine.

A Bundle can hold project specifications, decisions, procedures, tasks, patterns, or other Markdown knowledge. Multiple agents can read the same published files, while changes arrive as proposals and require ratification by a person or a recorded standing rule.

How it works

The reference implementation reads the Bundle and exposes seven agent tools. The documented capabilities include search, item reading, link traversal, citation, composition, and remembering; the README also states that the local tool server matches the tools available on item pages and the composition surface.

The underlying format is file-based rather than service-based. It defines typed links, graph views, canonical machine files, schemas, an OWL 2 RL vocabulary, and conformance vectors. The agsc command-line tool builds and checks the format, while agsc mcp provides the local MCP entry point. Nodes can be published as static files, but the format does not define a database, protocol server, or reasoning service.

Composition selects items, closes over their links, and emits seven kinds of starting-harness files. That output is an input for an architect or agent runtime, not a running application, and no renderer for a specific runtime is included in version 1.0.

Setup and configuration

The packages are published as agentic-system-core on npm and PyPI, with version 1.0.0 identified in the README. Install with npm install agentic-system-core or pip install agentic-system-core. After installation, invoke agsc mcp with a Bundle to provide the local MCP interface.

The supplied material does not specify environment variables, client-specific configuration, authentication, or a remote deployment mode. The described operation is stdio-based and uses files in the Bundle as its source.

Tools and capabilities

• Search knowledge items in a Bundle.
• Read item content and metadata.
• Follow typed links between items.
• Cite items by address.
• Compose selected items into a starting harness.
• Remember information through governed proposals.
• Support shared, read-oriented access to published Bundle files.

Limitations and notes

The reference engine claims Level 3 conformance, but implementations should be checked against the repository's specification and test vectors. The system is not a hosted knowledge database or a general-purpose reasoning server. Federation is walked by a client, and at version 1.0 the multi-node walk is described as an engine library function rather than a command.

AgenticSystemCore MCP server users should also distinguish published knowledge from proposed changes: the governance model requires a person to ratify changes directly or through a standing rule recorded in the node configuration. The generated harness is a collection of starting files, not an automatically running agent system.