A described API, and a refusal you can read.

The dashboard is a client of this API and nothing more — there is no second, private surface behind it. Every route is described by an OpenAPI 3.1 document served from the API itself, so a client can be generated rather than transcribed.

POST /mcp

curl -X POST "$NETWEFT_API/mcp" \
  -H "Authorization: Bearer nw_<prefix>_<secret>" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Or x-api-key. A key is shown once; only its SHA-256 hash is stored.

The description is served, not published

GET /openapi.json returns an OpenAPI 3.1 document generated from the route definitions themselves rather than written beside them. A route whose request or response shape changes changes the document in the same deploy, because the document is the route's own schema — there is no second copy to fall behind.

Two ways to authenticate

The dashboard authenticates with a session cookie, created by signing in with Google, GitHub or an email address and a password. The MCP server at POST /mcp authenticates with an API key: a bearer token, narrowed to some of 6 scopes, looked up by its public prefix and compared in constant time against a stored hash. The other routes take only the session cookie.

A key is displayed once and never again. It can be given an expiry when it is created, and it can be revoked — which leaves the row in place for the audit log rather than deleting the evidence that it existed.

Money is refused, not invoiced

A paid operation reserves credits before the vendor is reached. A workspace that cannot cover it gets a 402 with the shortfall in the body, before anything is bought — so a client can say by how much it is short instead of discovering an overage later. There is no overage, because there is nothing to overrun.

HTTP/1.1 402 Payment Required

{
  "code": "insufficient_credits",
  "error": "This workspace cannot cover the call.",
  "available": 1240,
  "required": 2000
}

402 — nothing was bought, nothing was charged

Failures have names.

A failing response carries a human sentence in error and a stable machine code in code. The sentence is free to change; the code is the contract, and each one has exactly one HTTP status so a name never means two things.

insufficient_credits HTTP 402: The workspace cannot cover the call. Answered before the vendor is reached, so nothing was bought and nothing was charged. available and required come with it, so a client can say by how much.
rate_limited HTTP 429: The fixed window is full. A Retry-After header comes with it, so a client library backs off on a number rather than on a guess.
insufficient_role HTTP 403: The caller is in the workspace but not with a role that may do this. Distinct from not being a member at all: the fix is a role change, not an invitation.
background_unavailable HTTP 503: The background service that runs audits and scheduled checks is unreachable. Nothing was charged and nothing was lost — the run simply cannot start yet.

Rate limits

Every policy is a fixed window. A refused request carries Retry-After, so a client backs off on a number rather than on a guess.

Spending endpoints per user — Keyed per user rather than per workspace, so a team of ten is not throttled by one member's script.: 60 per 60s
API key requests per key — Applied at the MCP endpoint against the key's public prefix.: 120 per 60s
Trial workspaces per workspace — Tighter while a workspace is on trial, and back to the normal policy the moment it is not.: 10 per 60s

What an agent can read through it

The dashboard is a client. So is whatever you write.

The API key is created in the dashboard, shown once, and scoped before it is shown. Everything it can reach is listed on the MCP page.

Start a trial See the whole rate card

There is no free tier. Three plans from $50 a month, each opening with a free trial.