agentshield.ogbuilds.aia human in the loop for your ai agents
agent·shield is a transparent http proxy for ai agents. safe traffic forwards instantly. destructive calls (a delete, a drop table, an rm -rf) are held in an approval queue for a human to approve or deny. change one url, you're done.
agent·shieldpending review
1
approved
312
denied
7
intercepted
1,284
body drop table customers
your agent is trying to wipe every customer record. agent.shield is holding it until a human decides.
agentshield.ogbuilds.ai/p/prod-agent
3 policies · 14ms median overhead
When a request looks destructive (a DELETE, a drop table, an rm -rf, a kubectl delete), agent·shield holds it in an approval queue and waits for a human to approve or deny. Then it forwards or blocks accordingly. Every decision, held or passed, lands in an audit log you can read later. Point your agent at one url and you have a human in the loop for anything that can't be undone.
agent·shieldbefore & after
an agent that can delete things
beforeone bad tool call runs a drop table with nothing to stop it
afterdestructive calls held in a queue for a human to approve or deny
adding guardrails
beforerewrite the agent or bolt on an sdk
afterchange one base url: no rewrite, no sdk
proving what the agent did
beforeno record of which actions ran
afterevery decision, held or passed, lands in an audit log
simple pricing
Free
put a human in the loop on one agent
1 agent proxy
- 100 intercepted requests / mo
- manual approve/deny queue
- 3 custom policies, 7-day audit retention
Pro
every agent, your whole team in the loop
10,000 intercepted requests / month (free: 100)
everything in free, plus:
- unlimited agent proxies
- ai risk scoring on every held request
- slack + webhook alerts
- team review + roles
- csv/json audit export
- 90-day audit retention (free: 7 days)
- unlimited custom policies (free: 3)
every plan includes
- transparent proxy: change one url, no rewrite
- regex policies on method, path + body
- destructive actions held for human approval
- every decision logged
a look inside
a few of the screens you'll actually use.
- drop-in proxy: change one url
- holds destructive calls for approval
- every decision logged
agent·shieldpending review
1
approved
312
denied
7
intercepted
1,284
body drop table customers
your agent is trying to wipe every customer record. agent.shield is holding it until a human decides.
agentshield.ogbuilds.ai/p/prod-agent
3 policies · 14ms median overhead
agent·shieldpending review
1
approved
312
denied
7
intercepted
1,284
body drop table customers
your agent is trying to wipe every customer record. agent.shield is holding it until a human decides.
agentshield.ogbuilds.ai/p/prod-agent
3 policies · 14ms median overhead
agent·shield guides
Ways to use agent·shield, and how it compares.
- use caseAI agent guardrails: stop the agent at the dangerous request, not after itAI agent guardrails should stop a destructive action before it lands, not log it after. agent·shield holds DELETEs, drop-table and rm -rf for human approval.
- how toHow to put a human in the loop for an AI agent's destructive actionsAdd human approval to an AI agent without rewriting it. Route its traffic through agent·shield, write hold policies, then approve or deny in a queue.
- how toApprove AI agent actions: a manual checkpoint for the calls that matterApprove AI agent actions before they execute. agent·shield holds risky requests in a queue so you can read the request and the matched policy first.
- use caseYour AI agent can delete production data. Here's how to make that safe.An AI agent with write access can delete production data on one bad inference. agent·shield holds DELETEs, drop-table and kubectl delete for approval.
- use caseAI agent firewall: the category, what it controls, and how agent·shield does itAn AI agent firewall sits between an agent and the systems it calls, inspecting every request and holding destructive ones. What to expect, with agent·shield.
- use caseMCP server security: put a gate in front of the tools your agent can callMCP servers give agents real tool access, destructive ones included. agent·shield proxies the traffic and holds dangerous tool calls for human approval.