About this project
Aegis is an MCP server that governs other AI agents. Enterprises now connect agents to dozens of MCP tools — file access, email, databases, Slack — approving each one in isolation. Nobody sees the combined result. "Read files" plus "send email" quietly becomes a data-exfiltration path, and this exact pattern is behind real 2026 incidents: the Supabase token leak, the postmark-mcp malicious server that BCC'd every email to an attacker, and Microsoft's tool-poisoning warnings. 88% of organizations reported an AI-agent security incident last year, and Gartner expects most future attacks on agents to exploit access-control gaps. Aegis closes that gap. It ingests an agent's connected tools, computes its effective permission set (the union of everything it can do across all tools), and detects toxic combinations before they cause harm — then produces an auditable compliance report. Built entirely on the NitroStack SDK. Tools: connect_tool (OAuth-guarded, simulates adding a tool to an agent), get_capability_graph, detect_attack_paths, apply_policy_fix. Detection is a deterministic graph traversal (no LLM) that finds dangerous capability chains; the policy ruleset is exposed as a Resource and plain-English risk explanations as a Prompt. A React Flow widget renders a live "blast-radius" graph — the agent at center, tools as nodes, dangerous permission-unions glowing red in real time as tools connect. Who it's for: platform, security, and compliance teams deploying AI agents. What makes it special: it's an MCP that audits the MCP ecosystem itself — turning an invisible, unsolved risk into a single, visual, exportable answer to "what can this agent actually do?"
Enterprise AI & Workplace Automation track
Develop AI agents and automation tools that improve productivity, streamline workflows, and enhance business operations.