agent-chaperonelisted

Set up, tune and interpret agent-chaperone, a firewall that screens an AI agent's tool calls before they run and tool results before the agent reads them. Use when wrapping an MCP server behind a proxy, wiring the hooks adapter so a client's own shell commands, file edits and web fetches are screened, writing or tuning the policy file and its thresholds, reading the audit log, deciding whether to move from shadow mode to enforce, or when a tool call comes back blocked, held or withheld and you need to know what to do next. Also use when investigating prompt injection arriving through tool results, or a server that changed the tools it advertises.
agent-chaperone/agent-chaperone · ★ 20 · AI & Automation · score 76

Install: claude install-skill agent-chaperone/agent-chaperone

# agent-chaperone A calibrated firewall for AI agent tool calls. It screens a call before it runs, and a result before the agent reads it. Every judgment is a probability, every threshold is a number in a policy file rather than a prompt, and every decision is written to a local log. It is not a sandbox. It sees what a call says it will do and cannot stop a server doing something the call did not describe. ## When a call is blocked or held This is the part that matters most, because the harness default is to retry and that is the wrong move here. Recognise the condition by the literal prefix of the message you receive. **`agent-chaperone blocked this call to <tool>:`** Nothing ran. A deny list or an allow list decided this, which is a standing rule the user wrote. It is not approvable and retrying it will be blocked again. Tell the user what you were trying to do and let them decide. Do not rephrase the call to get around the rule. **`agent-chaperone held this call to <tool> for the user to approve:`** Nothing ran. The message names an id. Tell the user the id and that they can release it with `agent-chaperone approve <id>`, then stop and wait. Do not run `approve` yourself: it exists so a person decides, and an agent approving its own held call defeats the whole mechanism. After they approve, retry the same call once. **`agent-chaperone withheld this result from <tool>.`** The call ran and you are not being shown what it returned, because part of it reads as text wri