agent-chaperone

At a glance

agent-chaperone screens tool calls and tool results in supported coding-agent integrations. It is worth starting in shadow mode if you want to see what a policy would flag before you let it hold calls or withhold suspicious output.

Good fit

Developers who control an agent’s MCP configuration or use a supported client hook and want to review tool traffic.

Consider something else when

Skip it if you need process isolation or must keep screened content away from the configured model backend.

Start here

Use a disposable configuration with a mock tool that returns harmless text written like an instruction. In shadow mode, check whether the log flags it and confirms the call was allowed. Compare against a normal response before considering enforcement.

Cost and requirements

Apache-2.0. The documented model screening uses a TypeSafe API key. Without the key, deterministic rules still run. Built-in tool coverage depends on client hooks, while the proxy covers configured MCP servers.

Limits to know

Shadow mode blocks nothing. Model screening sends arguments and results to the backend after pattern-based redaction. The proxy cannot constrain what a tool server does internally, and detection does not guarantee a block.

Related work

• Anthropic, maker of Claude Code, which supports the plugin (context)