Prevent dangerous calls

MCP guardrails: stop AI agents from deleting or leaking what matters

Set AI agent guardrails once: risky actions are blocked or wait for your approval, and everyday work keeps running.

Try for free Quick tour video

What Meandr's MCP guardrails stop your AI agents from doing

Each protection is one policy on a connected service. That policy covers every agent that uses the service, or only the agents you tag.

  1. 01 Deletion protection

    An AI agent can't delete your disks, databases or servers

    Add one deny rule in Policies, and no agent can delete or shut down your disks, databases and servers.

    • Nothing is deleted, and every other action on the service keeps working
    • Delete tools vanish from the agent's list, and any call that still tries is refused and shows up in Traffic

    For example: Railway, Supabase, Terraform, AWS, Kubernetes

    Real caseApril 2026: a coding agent deleted a company's production data and its backups in 9 seconds. March 2026: an agent wiped a team's servers and 2.5 years of data.

    Set up a deny rule in Policies

  2. 02 Sensitive data and access protection

    An AI agent can't read your secret data or change who has access without your approval

    Add an MCP human approval rule in Policies. The action waits while you see the agent, the tool and the full request, then you allow or deny it.

    • You get an alert in the app and by email
    • A deny stops the action, and your decision stays on record

    For example: Supabase, AWS

    Real caseResearchers used a fake support ticket to trick an AI agent into reading a database's secret keys and posting them back.

    How approval works

  3. 03 Secret files protection

    An AI agent can't write into files that hold your keys and passwords

    Add one deny rule in Policies for key and password files. The agent still edits the rest of your files.

    • The agent is told no, and the file stays as it was

    For example: a filesystem MCP server

    Real caseA flaw let agents write outside their allowed folder.

    Set up a deny rule in Policies

  4. 04 Refund protection

    An AI agent can't make a Stripe refund above your limit without your approval

    Set the limit in an approval rule in Policies. A bigger refund waits until someone you choose approves it.

    • Smaller refunds go through on their own
    • A denied refund never reaches Stripe

    For example: Stripe

    How approval works Stripe refunds with approval

Human in the loop for AI agents: questions

What happens when an MCP tool call requires approval?
Meandr pauses it, notifies the approvers in the app and by email, and runs it only if someone says yes before the deadline.
Does it work with any AI app?
Yes: Claude, ChatGPT, Cursor, Codex, VS Code or any other MCP client. If the app can't show the approval, the agent is told the action is waiting, and you decide in Meandr.
Can a tricked agent get past a rule?
A rule holds whatever the agent was told. Meandr checks every action before it reaches your service.
Why not just make the database read-only?
Read-only blocks every change. With Meandr, everyday changes keep working and only the risky ones stop.

Decide what your agents may never do

Connect your services through Meandr, then choose which actions are blocked and which wait for you.

Try for free

Questions? Email us