Overview
Properly securing agentic workflows requires a clear understanding of safety and security offerings at every layer of the stack. Stoplight's hook based approach positions it closest to where your agents actually leverage tools that transact with external systems.
Hook-level interception of agent tool calls at execution time.
Runtime tool control
Stoplight sits directly in the execution path for agent tool calls and decides whether risky actions should be allowed, denied, or escalated.
It evaluates each call against deterministic rules and AI policies before the action runs, providing real-time governance over commands, file operations, and MCP actions across every supported runtime.
Artifact and skill scanning
Scanning tools analyze code artifacts, dependencies, and agent skill definitions to catch known vulnerabilities, malicious payloads, or risky packages before they ever reach a runtime. This layer is valuable for supply-chain integrity and pre-deployment checks where patterns are well-understood and signatures are available.
Securing tool use
Well-intentioned instructions can lead to irreversible outcomes when an agent has unrestricted access to tools, because intermediary steps & failures can rarely be predicted.
“”
Observability First
Guardrail generation and maintenance with Stoplight is entirely automated and hands-free. Rather than asking teams to manually define every rule up front, Stoplight uses hooks to continuously analyze the tool calls your agents actually make, mines that usage data for emergent risk surfaces, and recommends guardrails in real time – automatically updating them as agents are equipped with new tools across any corner of your organization.
Stoplight supports a wide variety of agent runtimes, and we're constantly adding support for new ones, so you can secure your business against agentic tool calling risks across every provider your teams use from one place.
The only viable approach
Agents today operate across unbounded tool surfaces that grow constantly – new MCP servers, new CLI tools, new internal APIs – and no static ruleset, periodic audit, or manual review process can ever meaningfully keep pace. The only methodology that can sufficiently cover this problem is one rooted in continuous observation of what agents are actually doing, followed by automatic adaptation. This is not a product opinion; it is a structural constraint of the problem itself.