Your AI agent reads your Jira, Slack and inbox. Some of that text is talking to it, not you.
Description and use
mcp-warden sanitizes incoming data from external sources, removing hidden instructions and quarantining suspicious fields, enabling safe use of AI agents.
Problem
AI agents can ingest untrusted external text that contains hidden or explicit instructions aimed at the agent, potentially leading to malicious actions.
Outcome
mcp-warden effectively removes hidden instructions and quarantines suspicious fields, allowing developers to safely use AI agents with external data sources.
Workflow
- Install mcp-warden via brew or script
- Run
mcp-warden installto register hooks in supported agents - Configure trusted/untrusted servers in ~/.mcp-warden/config.toml
- Optionally wrap servers with proxy using JSON config
- Agent reads sanitized output; withheld fields trigger a notice with quarantine ID
- User reviews withheld content with
mcp-warden show ID - User releases content with
mcp-warden release IDor approves sensitive calls withmcp-warden approve TOKEN
Evidence
mcp-warden is one Rust binary. No service, no account, no network calls of its own. It works with Claude Code, Codex and Cursor on Linux, macOS and Windows.
It strips content a person would not see, withholds and quarantines fields that look like injection (the agent gets a notice instead of the payload)
The hook denies it, because an injected instruction could tell it to.
- Record date: 2026-10-12T14:41:18.095Z