Agent Control
Agent Control is the open source control plane for AI agents, built to establish a new standard for governing agent behavior at scale. It solves the problem of scattered, hardcoded checks by giving teams a centralized governance layer with step-level enforcement that can be managed from a single control plane and updated in real time without touching agent code. Developers can make any function governable by adding the control() decorator, turning meaningful decision points inside an agent into independently governed control points with their own policies. When a decorated function executes, Agent Control evaluates the input or output against the active policy and returns a decision: deny, steer, warn, log, or allow. If the decision is denied, the SDK raises a ControlViolationError before the unsafe action can proceed. Policies are decoupled from code, so developers decide where to place control hooks while policy teams decide what those hooks enforce.
Learn more
Preloop
Preloop is the open source AI agent control plane for agents that take real actions. It combines an MCP firewall for tool access, an AI model gateway for cost, safety, and attribution, policy-as-code with human approvals, runtime session observability, and audit trails in a single self-hostable platform. AI agents can deploy code, change infrastructure, move money, touch production data, and burn model spend in seconds, so Preloop helps teams control what agents can do, how much they spend, and which actions require human approval. It works with OpenClaw, Hermes, Claude Code, Codex CLI, Cursor, Gemini CLI, Windsurf, Cline, OpenCode, and any MCP-compatible agent or managed runtime. Access rules can inspect arguments and context, not just tool names, with CEL expressions for fine-grained conditions. Teams can start with observability, then layer in approvals and deny rules without SDKs or invasive app changes.
Learn more
Maetra
Maetra is an AI governance and compliance control plane for teams operating tool-using AI agents. Discover inventories agents and capabilities; Comply maps systems to applicable frameworks and keeps reusable evidence current; Govern evaluates consequential actions against versioned policies and routes human approval when required. Secure scans prompts, messages, model outputs, and tool calls for prompt injection, data exposure, unsafe actions, and policy violations. Task Guard detects task drift, scope changes, and mismatched effects. Interaction Guard protects supported browser-AI prompts and files, while Audit preserves linked decision, approval, runtime, and change evidence. Teams can adopt modules separately or together through the web app, REST APIs, SDKs, and MCP. A 14-day no-card trial is available, with paid plans from $20/month.
Learn more
HOL Guard
HOL Guard is a local-first runtime security layer for AI agents that watches what an AI assistant is about to do and stops risky actions before they happen. It sits between the agent and the computer, evaluating supported tool calls and local artifacts for threats such as secret and credential exposure, destructive commands, prompt-injection-driven actions, malicious or changed packages, risky MCP configuration, and unsafe plugins, skills, hooks, and settings. Known threats can be blocked automatically, while ambiguous actions are paused for user approval so people remain in control. Guard runs entirely on the developer’s machine, works offline, and does not upload files, prompts, or passwords. Local checks typically complete in under 50 milliseconds and require no changes to existing code or routines. It supports coding agents including Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Hermes, and OpenClaw, with tailored integrations that inspect actions before execution.
Learn more
