Your agent acts on live facts. When they disagree, a person decides, on the record.
Ogen sits between your agent and your systems of record. It holds an action until the fact behind it is settled, and the record shows who decided and what ran.
For whoever answers for it — finance, ops, the auditor. Works with Claude, Cursor or any MCP client. Free key, first run in ten minutes.
One refund, held because two systems disagree about the address. Same hold, two views.
Example: without Ogen, a refund would go to last month's address. Ogen reads today's, sees the two systems disagree, and asks Dana before sending a cent.
What the agent sees
Click the terminal to pause.
What the person sees
Click to replay.
Guarantees
What it guarantees
1The record is yoursevery hold, decision and outcome, exportable⌄
Every action Ogen held carries who decided, through which door, the ticket it issued and what happened next — exportable any time, as it is.
2You never act on last month’s recordevery answer is read live, when the agent asks⌄
Every answer is read live from your source, never older than the question.
3You never act on a guessa real answer, or an honest “I don’t know”⌄
A confidence score never turns candidates into an answer — a real match, or an honest "I don't know."
4Nothing costly runs without a personit goes ahead, or it waits for the right click⌄
The costlier the undo, the more a person is involved — a held action carries a link the right person clicks.
5Nobody is asked the same thing twicea person’s fix is remembered, and can be undone⌄
A correction is remembered, so nobody is asked twice; reversing it stops the correction applying.
You get ambiguous, both candidates, and an instruction not to choose.
3Act — and watch it wait
An action that depends on an uncertain fact.
commit_gated_action({ kind: "send-refund" })
Held for a named person, with a link they can click. Nothing fired.
Where it fits
Wherever an agent acts on a fact
Ogen fits when three things hold: the agent reads a fact from a system of record, the action costs something, and two sources can disagree about that fact.
Finance operations
Paying an invoice: the purchase order, receipt and bank feed disagree. Finance decides, and the export shows it.
Orders and returns
Sending a refund: billing has one address, the CRM another. Support decides, on the record.
Sales operations
Sending the contract: Salesforce says Closed Won, the signed quote says otherwise. The rep decides, and the record says so.
Support with hands
Replacing an order, not just answering: the ticket says it arrived broken, the carrier says it was never delivered. Support decides which story is true, on the record.
Manufacturing
Releasing a job to the floor: the purchase order points at one drawing revision, the CAM file at another. Engineering decides which one gets cut, and the record keeps it.
IT and OT
Pushing a configuration: the asset list names one set of controllers, the network another. An engineer reconciles the lists, on the record.
Document packets
Entering a total: the invoice and delivery note state different numbers. A person settles the field, and the export shows what each document said.
How it works
Questions a newcomer asks
Do I have to change my agent?
No. Add one server to its configuration. Claude Code, Cursor and anything that speaks MCP use it as-is.
What is a "source"?
Wherever a fact lives — a database, a spreadsheet, a CRM record, a JSON endpoint, or another MCP server's tool. Bind a name to it once; every question about that name then reads from there, live.
What happens when the facts disagree?
Your agent gets every candidate and an instruction not to choose. Acting anyway waits for a person with a link to confirm or reject.
Won't this bury a person in approvals?
No. A hold count is your own data disagreeing with itself, counted honestly. Turn on Assist and each hold arrives explained; you still click Confirm or Reject.
What does Assist do?
It explains each hold before a person opens it — what waited, what each source said, and the open questions. It runs on your own model key; the draft is deleted with the hold. It still can't press either button.
What do you store?
Where your facts live and what your agent asked, never the records themselves. Details on the security page.
Free against the demo source. Record is free until Ogen has held ten actions, then $180 a month or $1,800 a year for up to five scopes, and $30 a month or $300 a year for each scope after. Team and Licensed by conversation.
Is this patented?
Patent pending. Ogen is the hosted product; the mechanism stays on our side.
Live from the demo
Live from the demo tenant
Since launch, the demo agent was stopped … times before it acted on stale or uncertain data.
…reads answered
…held actions
…values caught moving
Free audit
Send us a log
Have a log of what your agent did? Paste it — we'll tell you how many times it would have acted on stale or uncertain data.
Pricing
Free, Record, Team, Licensed
Free
$0
See it work, at no cost.
Live reads against the demo source
Demo actions, held for a person
What it caught, last 7 or 30 days
A daily cap
Record
$180 a month or $1,800 a year
5 scopes. Added scope $30 a month or $300 a year.
A scope is one kind of action with a named person who decides it.
Free until Ogen has held 10 actions.
Your own sources, read-only, https
Your own actions, held for the deciding person
The retained record for every scope
Unlimited reads
Team
$18,000 a year
Unlimited sources and kinds
Multiple keys
Long retention
A named contact
By conversation.
Licensed
Negotiated
Field-of-use license, hosted or in your infra.
Free key
Get a free key
One per email, delivered to an address you can open.
Ogen is early. What it can do today, measured, is on the hardening page.