Guides

Four things people set this up to do. Each one says what it covers, what the configuration looks like, and where it stops, which is the part most of this kind of writing leaves out.

They assume the package is installed and describe the current release. If you want the architecture rather than the setup, that lives in the design document.

npm install -g agent-chaperone

The guides #

• Screen MCP tool calls before they run

  One change to a client configuration puts a server behind the screen. What crosses the proxy, what the pre-call and post-result screens ask, and where a proxy stops seeing anything.

• Screen tool results for prompt injection before the agent reads them

  A fetched page, a file and an API response are all text somebody else wrote, and the agent reads them as input. What the screen catches, and the measured count of what it does not.

• Screen Claude Code's built-in tools

  A proxy sees MCP traffic and nothing else. Three hooks put the same screens in front of the shell, the file edits and the fetches a client runs itself.

• Catch secret-leaking tool calls before they run

  Reading a file and posting to a URL are both ordinary. The pair is the problem, and a list of allowed tools cannot express it.

Everything else #

• Measured results

  One dated run against public prompt-injection benchmarks and hand-labeled tool calls, with every threshold, what was caught, what was missed and what was flagged in error.

• Design

  The architecture, the exact wording of every screening question, the policy file and the audit log format.

• Decision records

  Why the main choices went the way they did, including the ones that were reversed.