Your agent creates the connection. You are the only one who sees the key.

Connect Enroutia to Claude, Cursor or n8n's MCP node and ask it, in one sentence, to set the connection up: name, model and budget. Five minutes, and the configuration step stops existing.

The sentence

With the server connected, this is literally what you type:

“Connect my n8n automation, step 1, to Mistral through Enroutia. Cap it at €20 a month.”

1 · Create an account token

Open the panel, under Settings.
Under “Account tokens”, name it after the machine that will hold it (“Claude on my laptop”) and pick its permissions.
Copy it: like an API key, it is shown exactly once.

The two permissions are independent, not a ladder: “Read the account” sees balance, usage and the list of connections; “Manage connections” creates, rotates and revokes keys. A token with the second and not the first is a perfectly sensible thing to grant.

The quick way: “Add connector”

If your application knows how to add MCP servers on its own — Claude on desktop and on the web, among others — give it just the URL below. The server supports what those applications ask for: it registers them dynamically and shows a screen where you choose what to grant, with no token to copy. We have not yet watched that flow end to end inside one of those applications, so if it stalls, the steps below always work.

The steps that follow are for everything else: n8n, Cursor, and anything that wants a header in a config file.

2 · Point your agent at it

It is a remote MCP server over HTTP. Wherever your tool asks for a URL and a header:

URL: https://api.enroutia.com/v1/mcp
Transport: HTTP
Header: Authorization: Bearer pat_live_…

In Claude Code or Claude Desktop, in your MCP configuration file:

{
  "mcpServers": {
    "enroutia": {
      "type": "http",
      "url": "https://api.enroutia.com/v1/mcp",
      "headers": { "Authorization": "Bearer pat_live_…" }
    }
  }
}

3 · What it can do

Seven tools. Your agent picks the right one; you speak English.

list_models — the catalogue, with the price of 1.000 messages for each model.
get_balance — credit left, and how many connections exist.
get_usage — spend per day, up to 90 days back.
list_connections — your active connections with their model, cap and spend. Never their secrets.
create_connection — create a new connection, optionally pinned to one model.
rotate_connection — replace the secret of a connection that has leaked.
revoke_connection — revoke one connection. One, never several.
chat — send a message to a model and get the answer, billed to your credit.
compare_models — the same prompt against every model, with what each answered and cost.

The key never comes back through the agent

When your agent creates or rotates a connection it does not receive the key. It receives a single-use URL, good for 15 minutes. You open it, already signed into the panel, and copy the key once.

That is one extra step, on purpose. Anything returned to a model is kept in the transcript of that conversation — by whoever runs the model, not by us — and replayed into every later turn. A key in there is a key in a third party's logs for as long as they keep them, and no revocation of ours reaches it.

The permission that spends money

“Run models” is a separate permission, and neither of the other two implies it: “let my agent set up my connections” and “let my agent spend my balance” are different sentences, and somebody who meant the first should not have granted the second by accident. Leave it unticked and your agent can create and rotate connections but cannot spend a cent.

What it deliberately does not do

Every agent-created connection carries a mandatory monthly cap. A person can decide they want no cap and be right; a program acting on a sentence cannot make that call for them.
Rotating and revoking ask for the connection's exact name, not a “yes”. Typing the name fails precisely when the model has the wrong connection in mind, which is the mistake worth catching.
It does not call the models unless you grant “Run models”, which is unticked by default.
It cannot touch billing, cannot top up credit, and cannot delete your account or mint more tokens.