Clyro

The core of Clyro is clyro.wrap(). You pass your agent plus a config object, and it returns a governed version that traces execution, tracks cost and enforces limits without changes to your agent logic. That last part is the real selling point. Retrofitting guardrails into a working agent is usually painful, and Clyro's design keeps the change to a handful of lines.

Clyro groups its controls into a four-part stack: step limits, cost limits, loop detection and business logic guardrails. Step limits cap how many actions a run can take. Cost limits stop a run when spend crosses a threshold. Loop detection watches for repeated execution paths. Why does that matter? A stuck agent can burn a budget overnight. Guardrails check actions against rules before they fire. These are AI agent guardrails that act during a run, not after it, and they target failures observability tools only report once the damage is done.

Policies live in YAML, and each rule pairs a condition with a decision. You can block an action outright, allow it, or require human approval before it proceeds. Rules come with eight operators for matching fields, and each rule fails open, so a broken policy doesn't take your agent down with it. For teams that need a paper trail, that per-rule behavior is worth understanding before rollout.

Clyro ships five adapters: LangGraph, CrewAI, the Claude Agent SDK, the Anthropic SDK and a generic option for any Python callable. That spread matters because agent stacks fragment fast, and a governance tool that only supports one framework forces you to either switch or skip it. The generic adapter is the escape hatch for whatever you built by hand.

Beyond the SDK, Clyro ships an MCP wrapper that acts as a JSON-RPC proxy for Claude Desktop, Cursor and VS Code, and a set of Claude Code hooks for PreToolUse and PostToolUse events. The hooks govern Bash, Edit and Write, which means you can block something like DROP TABLE before the tool ever runs. This is a different surface than agent-run governance, and it broadens where the product applies.

Clyro runs fully offline with local YAML policies and needs no API key. The cloud is optional and adds shared dashboards, policy sync, drift detection, flame graphs and trace comparison as you move up the tiers. That local-first design suits teams with data policies that make sending execution traces to a third party a non-starter, and it means you can evaluate the tool without a signup.

The SDK calculates LLM costs automatically for OpenAI and Anthropic models, so the numbers behind your cost limits come from real usage rather than a guess. If your stack runs other providers, you should expect to supply cost data yourself, since the automatic calculation is scoped to those two families.