Reading an SEO workspace from an agent

Most SEO tools are built for a person with a browser. The data underneath them is perfectly readable by something else, and the something else is increasingly the thing asking.

Two surfaces, one product

netweft publishes both.

The HTTP API is described by an OpenAPI 3.1 document at /openapi.json. Everything the dashboard does goes through it, so a generated client is a real client rather than a partial one.

The MCP server sits at POST /mcp, authenticated by API key. It exposes a workspace’s audits, keywords, rankings and project overview as tools an agent can call directly, without you writing the glue.

Neither is a side project bolted on after launch. The API is what the dashboard talks to; MCP is the same data under a protocol an agent already speaks.

Reading is free, because looking was not

The credit model makes this cleaner than it usually is.

When netweft checks a keyword’s position, it pays a vendor for a SERP fetch and charges the workspace for it. That is the moment money changes hands. The observation is then stored.

Reading that stored observation later — from the dashboard, from the API, or from an agent — costs nothing. There is no second invoice, so there is no reason to meter the read. An agent can pull a project’s entire ranking history every hour and the balance does not move — which works only because that history is stored rather than recomputed.

The corollary matters too: an agent cannot run up a bill by reading. The operations that spend credits are the ones that go to a vendor, and they are the same operations, at the same prices, whichever surface asks for them.

What the assistant is allowed to say

netweft also ships its own AI: a per-project chat and an onboarding strategist. Both are restricted to figures read back through tools.

That constraint is not a disclaimer, it is an implementation. The model cannot answer “what is my average position for this cluster” from its own weights, because the number is not in its weights — it has to call a tool, and the tool returns what is stored. If the tool has nothing, the answer says so.

An SEO assistant that invents a plausible number is worse than no assistant, because a plausible number is exactly the kind of thing nobody checks.

A practical shape

The pattern that works, in our experience of building it:

• Give the agent an API key scoped to one workspace.
• Let it read rankings, audits and the project brief. These are cheap, stored, and safe to poll.
• Keep the operations that spend money — a fresh SERP fetch, a new crawl, a backlink pull — behind an explicit decision, whether that decision is a human pressing something or an agent with a budget you set.

The credit ledger is what makes the second half tractable. Every spend has a line, a reservation and a settlement, so an agent’s behaviour is auditable in the same units as everything else.