Repolode
Submit
Justin0504

Justin0504/Aegis

Runtime firewall for AI agents that intercepts, classifies and blocks tool calls before they execute, with approvals and a tamper-evident audit trail.

View on GitHubWebsite
Stars
523
Forks
55
Open issues
5
Last commit
8 days ago

What is it?

AEGIS sits between an AI agent and the tools it calls. Each tool call is classified by kind (database, file, network, shell, prompt injection, supply chain), checked against policies and anomaly baselines, and then allowed, held for human approval, or blocked before it runs. The project targets cases such as destructive SQL, unexpected data exfiltration, or hallucinated shell commands produced by an agent.

Integration is done through an SDK that wraps Python, TypeScript and Go LLM clients, or through HTTP and MCP stdio proxies for agents that cannot be modified. The gateway runs as a Docker service with a Compliance Cockpit dashboard for live traces, policy editing, approvals and a kill switch. Policies can be written as AJV JSON-Schema rules or a per-tenant YAML DSL that can only tighten decisions.

Every allowed call is recorded in a SHA-256 hash-chained trace log, with optional Ed25519 signatures and an RFC 6962 transparency log for offline verification. The project is MIT-licensed and includes compliance bundles, SIEM and OTLP export, and a published agent threat ontology.

Key features

  • 01Classifies each tool call by kind before policy evaluation, without requiring tool annotations.
  • 02Holds high-risk calls for human approval in the Cockpit before they execute.
  • 03Ships default policies for SQL injection, file access, network access and prompt injection.
  • 04Adds a per-tenant YAML policy DSL that can only tighten decisions.
  • 05Records traces in a hash-chained, optionally Ed25519-signed audit log.
  • 06Works through Python, TypeScript and Go SDKs, HTTP proxy and MCP proxy.

Use cases

  • Putting guardrails on agents that can run SQL, shell commands or file operations.
  • Requiring human sign-off before an agent publishes packages or deploys to Kubernetes.
  • Auditing agent behavior for compliance reviews with tamper-evident logs.
  • Controlling closed-source or compiled agents by routing their LLM and MCP traffic through a proxy.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/Justin0504/Aegis
  2. 2

    Change into the project directory

    cd Aegis
  3. 3

    Start the services with Docker Compose

    docker compose up -d
  4. 4

    Enable AEGIS in your Python agent by importing it once

    python -c "import agentguard; agentguard.auto('http://localhost:8080', agent_id='my-agent')"

Things to know

  • The desktop app download is currently for macOS on Apple Silicon only, and Windows support is not yet available (WSL2 is suggested).
  • The manual setup requires Docker and running several services locally.

In collections

Claude Code tools, skills & pluginsOpen-source AI agent frameworks & appsMCP servers

Topics

#ai-agents#agent-security#ai-safety#audit-trail#policy-engine#mcp#llm-observability#anthropic#arxiv#langchain

FAQ

What is Aegis?
Runtime firewall for AI agents that intercepts, classifies and blocks tool calls before they execute, with approvals and a tamper-evident audit trail. AEGIS sits between an AI agent and the tools it calls. Each tool call is classified by kind (database, file, network, shell, prompt injection, supply chain), checked against policies and anomaly baselines, and then allowed, held for human approval, or blocked before it runs. The project targets cases such as destructive SQL, unexpected data exfiltration, or hallucinated shell commands produced by an agent.
Is Aegis free to use? What license does it have?
Aegis is open source under the MIT license. Check the license terms before using it in commercial products.
Is Aegis still maintained?
Actively maintained. The last commit to Aegis was on Oct 31, 2026.
What can I use Aegis for?
Putting guardrails on agents that can run SQL, shell commands or file operations. Requiring human sign-off before an agent publishes packages or deploys to Kubernetes. Auditing agent behavior for compliance reviews with tamper-evident logs. Controlling closed-source or compiled agents by routing their LLM and MCP traffic through a proxy.
How do I install Aegis?
Clone the repository: git clone https://github.com/Justin0504/Aegis Change into the project directory: cd Aegis Start the services with Docker Compose: docker compose up -d Enable AEGIS in your Python agent by importing it once: python -c "import agentguard; agentguard.auto('http://localhost:8080', agent_id='my-agent')"
Where can I get Aegis?
The source code and installation instructions are on GitHub: https://github.com/Justin0504/Aegis

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects