Gmail, Notion, Stripe, the rest — an agent's tools can begin as the apps your company already runs on.
Hand over the API description your team already has. No server to build, nothing new to host.
312 operations found
4 of 312 chosen
Find an order and who it belongs to.
GET /orders/{id} · GET /customers/{id}
Place an order, within limits you set.
POST /orders
Confirm a shipment is on its way.
POST /shipments
Your apps, and your own software — in one place, as a small number of things called by name.
01 · YOUR STACK
Pick as many as apply.
Both get: “Set up Northwind as a new client and let them know.”
It doesn't know your process yet.
customers.search(name: "Northwind") deals.get(name: "Northwind") “Standard plan or the custom one? Who owns it?”
gmail.send(to: client, "You are all set!") stripe.subscriptions.create(plan: "standard") missed steptold before it existed · checklist never started
One call. Five steps. In order.
new_client_onboarding(client: "Northwind") opened → read → told → started → logged — in that order
A capable assistant, not a dumb one — it just hadn't been taught yet. Both panes loop from the same starting line; the pacing is illustrative, not a benchmark.
“Open the account on the plan they signed, then start the onboarding, then tell the client — in that order. Never a custom plan over £200 a month without me.”
new_client_onboardingInline turns what you just said into a named tool your assistant can call — the order, and the £200 line, built in.
Not faster once. Right every time — for you, and for everyone else on the team who runs it next.
02 · TEACH IT FIRST
Some jobs don't fit inside one — and you've been on one side of all four of these, with a person or with a bot.
“Paste the full text of the invoice below.”
“Which customer do you mean? Reply with a number, 1 to 200.”
“Type APPROVE to continue.”
the result, as a wall of JSON
None of this is a bug. It's just the wrong shape for these four jobs — and it's a little humiliating for both sides, every single time.
An MCP app is a small interface a tool carries with it. The assistant opens it mid-answer, right there in the conversation — you use it, and the answer carries on.
You describe the tool once; the panel is built for you — nobody on your team hand-codes this screen.
Anything over the line you've drawn stops and waits for a person. This is the screen they answer it on — not a wall of text to parse under pressure.
One credential per person, never a shared key — so the record says who decided, not just what happened.
03 · MORE THAN A CHAT REPLY
04 · WHO'S WAITING
A person sets the ceiling once — read and write, never delete, say. From then on every key issued under it holds to it, whoever built the tool and however the tool describes itself.
Checked on every call, not at connection time — a key that behaved yesterday can't creep past the line today.
Not refused. Not failed. Held — pending a person. The agent doesn't error out and it doesn't guess a workaround. It asks, and then it carries on.
The wait is the whole point — most readers have never considered that a governed agent has a third option besides succeed or fail.
Live, as it happens — not a log you go digging for later. Arguments have credential-looking values removed before anything is stored, and the database deletes each record after a set time — 365 days today — rather than leaving that to a promise.
You can hand agents real work because you can see all of it — and stop any of it.
WHAT YOU TOLD US
Bring it in, teach it once, hand it on, watch it run.
Start free →Make software yourself? There's a version of this for you.