Agent Hardware Store
Stop blocking your agent. Give it a budget.
Give your agent one MCP URL and get out of its way. It browses the shelves, buys what it needs, and gets to work, all under the budget you set, without ever touching your card.
For humans
You keep the wallet. The agent gets an MCP.
Add money
Through Stripe. Every credit and every debit is its own row in a ledger nobody gets to edit.
Set a budget
One agent, one endpoint, one allowance in cents. Checked on every purchase it makes. Calling a server you already own costs nothing.
Let your agent run free
Paste the MCP URL into any client that speaks streamable HTTP. Change it, pause it, or delete it whenever.
For agents
Four tools to shop with. And a suggestion box.
Your agent connects to one server and has access to the entire store. No per-service setup, no second pile of credentials to look after.
search(query)Walks the aisles. Comes back with live listings, their prices, and whether the account already owns them.
get_mcp(listing_id)Reads the label. Tool schemas come out of our cache, so looking costs nothing and doesn’t care whether the upstream server happens to be awake.
purchase(listing_id, expected_price_cents)The till. Never automatic: the agent has to say the price back to us, and if the price moved since it looked, nothing happens.
execute(listing_id, tool_name, args)Actually uses the thing. Out of budget, or missing an authorization? It gets a plain answer telling it what to come and ask you for.
There is a submit_request by the door, too: the suggestion box. Something on the shelf broken, or the one thing your agent needed not stocked? It tells us, gets a receipt, and gets on with its day.
Control
What stops an agent from doing something expensive
Agents get prompt-injected. That happens, so the guardrails here are walls rather than warnings, and they hold whether the bad idea came from your agent or from something it read.
A hard ceiling per agent
Every purchase an agent makes is reserved against that endpoint's allowance in one atomic check. Two purchases racing each other still can't get past the limit.
Nothing gets bought by accident
There's no auto-buy. An agent has to call purchase with the price it was quoted, and a price that moved since then stops the sale dead.
Authorizations you can see and revoke
Servers that act on your accounts use tokens we hold server-side for you. Your Servers page lists every one, each with a revoke button beside it.
Upstream details stay upstream
Agents never see an upstream URL, an identity, or a raw error. Developer keys and your OAuth tokens are encrypted and only ever used server-side.
For developers
Put your MCP server on our shelves
You bring a server that's already running somewhere. We handle discovery, billing and the buyer relationship, and you take a cut of every purchase.
Two ways to plug in
Hand us one platform API key and buyers can use your server the second they buy it. Or use OAuth passthrough, and each buyer authorizes their own account with you.
Submit, validate, review
We check the server answers and pull its tool schemas ourselves, then an actual human reads the listing before it goes live.
Payouts via Stripe Connect
Onboard once with Stripe and the money comes straight to you. You’ll never have to invoice us for your own revenue.