MCP Agents

Public MCP and agent workflow experiments focused on tool routing, safe automation, and inspectable agent behavior.

This repo presents a small Model Context Protocol server surface for Primus OS-style agent workflows. The portfolio story is not autonomous execution for its own sake; it is controlled tool access, deterministic constraints, and operator-visible behavior.

What This Demonstrates

• MCP server structure for exposing agent-accessible tools.
• Tool routing for document audit, operations, and decision-support workflows.
• A modular execution layer that can be reviewed and constrained.
• Public-safe framing for agent/tool interfaces.
• The control pattern: ACTION -> VALIDATE -> APPROVE -> EXECUTE -> LOG -> REPLAY.

What This Does Not Claim

• It is not a production-grade agent platform.
• It does not grant agents unrestricted authority.
• It does not include private memory, credentials, customer records, or production runtime logs.
• It does not prove compliance or financial safety by itself.

Architecture

• server.js - kernel MCP server.
• os/primus.os.js - operating rules and constraints.
• os/council.config.js - council-style decision configuration.
• tools/*.js - MCP tools for audit, quant, and operations workflows.
• memory/ - local operator context; keep real memory private.

Safe Demo Path

npm install
node server.js

Use synthetic prompts and sample inputs only. Do not connect this public demo to private systems, secrets, production tools, wallets, customer files, or live execution surfaces.

Public/Private Boundary

Public code should show the tool interface and governance pattern. Private deployments should keep real operator memory, runtime journals, credentials, and production integrations outside the repo.

Topic Recommendations

mcp-agents, ai-governance, agent-safety, human-in-the-loop, tool-routing, workflow-automation, public-safe-demo