Making a product agent-ready with MCP

Why the agent surface, not another feature

A growing share of research work is done by AI assistants acting on a user’s behalf. An assistant cannot click through a dashboard — it needs tools: typed, documented functions it can call and whose answers it can cite. MCP is the emerging standard for exactly that, supported by Claude, Cursor and a fast-growing list of clients. For a product company the implication is blunt: your next user may be software, and it will choose whichever provider its tools can actually reach.

What we exposed

We did not build new functionality. The MCP server fronts what the product already had: SEC-filed statements back to 1985, DCF and fair-value models, filing and transcript search that returns the exact cited passage, supply-chain facts from 10-Ks. Each became a tool with a narrow contract — one job, typed inputs, an answer an agent can quote with its source. The same endpoints are also plain REST, so teams that are not on MCP yet lose nothing.

Pricing had to change shape

Subscriptions assume a human who logs in weekly and values access. An agent workload is spiky: nothing for days, then a burst of calls during one research task. So the agent surface is pay-as-you-go — a free key with 900 free calls to remove all friction from the first integration, then small credit packs from $10. That shape is what converted: the first paying customer bought credits for occasional agent calls, a purchase a monthly subscription would likely never have captured.

What agent-readiness required that the web app did not

Three things surprised us. Docs are for machines now: an OpenAPI spec and an llms.txt matter more than a polished features page, because the integrating developer often lets their assistant read them. Every answer needs a citation: an agent that cannot show where a number came from produces output its user cannot trust, so tools return the filing reference alongside the value. Discovery moves to registries: agent developers find servers in MCP directories the way humans find apps in app stores — being listed is distribution.

Is this right for your product?

If your product holds data or performs a computation that shows up in anyone’s AI-assisted workflow, an MCP server is a small, well-bounded project — weeks, not quarters, when the capabilities already exist behind an API. It is also a genuinely new sales channel with its own pricing logic and its own discovery surfaces. That combination — scoped engineering with a measurable commercial hypothesis — is the kind of AI transformation step we help Nordic companies take first.