How We Built Bassethound in 48 Hours

The challenge

We wanted to know how fast a small team could take a real product from an idea to something deployed and supported using agentic engineering. The answer had to come from a service we would have to run, not from a demo.

Bassethound was the test. Ask an AI agent about a company and it answers from memory: stale, uncited, and wrong often enough on the things that move that you cannot use it for a sales call or a vendor review. The underlying data exists, but it is split across vendors. One knows the tech stack, another firmographics, a third infrastructure, a fourth security. Each wants an account, a key and an integration, and none of them correlates with the others.

For an agent that is worse than for a person. Four integrations before the first useful answer is four reasons the capability never gets built at all.

What we built

Forty-eight hours, idea to deployed. The whole thing was built with agentic engineering: the service, the crawl and probe adapters, the detection rulesets, the identity and billing paths, and the deploy. Not a prototype that got rewritten afterwards. This is the service running today.

One call, five correlated layers. A domain goes in and a dossier comes back: technology stack, infrastructure, firmographics, AI-readiness and security posture, joined rather than listed. The stack layer is commoditised and we treat it that way, because every stack vendor now ships an MCP server. The layer worth having is AI-readiness: llms.txt, exposed MCP endpoints, model providers, vector stores, LLM SDKs. It tells you whether a company has shipped anything or is a year away. That did not exist over MCP anywhere else when we built it.

Everything is read live from the site, DNS and headers at the moment of the call, so it beats a model recalling a company from training data and it can say what it saw.

Signup inside the conversation. This is the part we are most pleased with. There is no website to visit, no form, no API key to paste into a config file. Identity runs through WorkOS AuthKit and checkout through Stripe Agentic Commerce, which is standard Stripe Checkout wrapped in an MCP pattern, so a user upgrades without leaving the chat they are already in.

The hard part of that is not the checkout. It is the moment after. We keep the entitlement server-side, keyed by the OAuth subject in the shared Redis the service already touches for rate limiting, so the instant the Stripe webhook lands the very next tool call is paid. No token refresh, no re-authorisation, no “log out and back in”. “I paid and I am still rate limited” is the worst possible moment to have a seam, and this design has none.

It also means the upgrade works on hosts that do not support elicitation at all, including claude.ai on the web and ChatGPT, because nothing depends on the client cooperating.

One subscription in place of four. Stack, firmographics, infrastructure and security were four vendors, four contracts and four integrations, and none of them joined up. This is one call against one bill, which is the cost argument as much as the convenience one.

Distribution without an integration. Bassethound is listed in the Claude directory under its own name. Adding it is a click rather than a project, which is the entire argument for building this as a remote MCP server rather than a SaaS product with an API somebody has to wire up.

Stateless by construction. Every call is self-contained, with no session and no initialisation handshake, so any replica serves any request and scaling out is free. Redis is the only state.

What changed

The build time is the finding. A hosted service with authentication, metered billing, live data collection and a public listing went from idea to deployed in 48 hours. That number is the reason this page exists, and the method behind it is the one we bring to client work.

Nobody integrates anything. Reaching it takes no key and no account. Paying for it takes no integration either.

The answer is checkable. It comes from the live site rather than a model’s memory, so it carries what it saw and when.

It qualifies our own pipeline. The AI-readiness signal that tells a user whether a prospect is ready is the same signal that tells us.

Why NimbleBrain

The interesting claim is not that we shipped fast. It is what we shipped fast: identity, metered billing, an entitlement with no refresh seam, live data collection and a directory listing, in 48 hours, all of it live.

That is the argument for agentic engineering as a delivery method rather than a demo. The work a team does in a quarter is mostly not typing, it is the loop between deciding, building and finding out. Compress the loop and the scope of what a small team can carry changes shape. Bassethound is the version of that claim we can point at, because it is running and we support it.