Keeprun stops your AI agents before they refund, pay, or message the wrong thing. One URL, no rewrite.
Drop-in MCP proxy or a 3-line SDK. Works with any MCP client.
Refund bot wants to run
Refund ₹18,400
Held: order total is ₹1,840.
The problem, in plain words
These are published benchmark results, not Keeprun results. A human verifies every citation before launch.
Tools fail silently
A timeout or empty result looks like a normal answer, so the agent carries on.
MLSys 2026, Measuring Agents in Production: across 86 systems, 74% rely on human checking.
Agents answer confidently wrong
The same request can land differently from one run to the next.
Consistency Gap (arXiv 2609.08832): 77% success per run, versus 53% when all 5 runs must succeed.
Some actions cannot be undone
A refund, a payout, or a message to a customer cannot be taken back.
PayBench: agents fell for trap payments 12 to 50% of the time; outside checks cut that to 5 to 11%.
How it works
Swap one URL
Point your agent at the Keeprun URL instead of your tool server. No rewrite.
Keeprun checks every call
It catches failures, diagnoses them, and holds risky actions.
You approve the risky ones
From the dashboard, email, Slack, Teams or WhatsApp, on your phone.
What it catches
Diagnose, don't just retry
Without Keeprun
refund(order_1042, 18400) -> error refund(order_1042, 18400) -> error refund(order_1042, 18400) -> error (agent loops, or gives up)
With Keeprun
[Keeprun] refund not executed. Refund 18400 is greater than order #1042 total 1840. Re-read the order and use its total, or ask the user. Do not retry.
On the ToolBench-X benchmark (Tian, Shi & Zhao, arXiv 2606.25819), a short diagnosis hint recovered roughly 60 to 80 percent of the lost accuracy, while extra attempts without a diagnosis added only 3.5 to 11.5 points. That is the idea Keeprun is built around.
Approvals anywhere
Approve or deny with one tap. The agent gets the result back.
Slack
Approve or deny with one tap. The agent gets the result back.
Teams
Approve or deny with one tap. The agent gets the result back.
Approve or deny with one tap. The agent gets the result back.
Works with
Any MCP client, the OpenAI Agents SDK, LangChain, n8n, and plain Python or TypeScript through the SDK.
Built by operators
We built Keeprun to protect our own agents at Platestory. We use it exactly as a paying customer would.
A case study will link here once the pilot has measured data.
Founding pilot
$49/mo locked for life, first 20 customers. 20 spots left
Growth features, direct founder line.
FAQ
Does it slow my agent?
Checks are designed to add little overhead, and the heavier LLM judge runs after the fact. Holds wait only when a rule asks for a human.
What if Keeprun is down?
By default calls fail open and are logged as unchecked, so your agent keeps working. You can set payments and refunds to fail closed.
Do you store my data?
Payloads are redacted in the gateway before anything is stored. Full-payload storage is opt-in per agent and encrypted.
Do I need a developer?
To swap the URL, yes, briefly. After that, approvals and rules are plain-English and work from your phone.
One URL. No rewrite. Start on the free plan.