Why is fusion better than fanning out to five MCPs?

Best move: ask one server for the whole picture, not five servers for five pieces you have to staple together.

Why it works: correlation only holds when the layers come from the same observation. One crawl feeds all five detectors, so the links between them are measured, not guessed.

Key takeaways

• Fan-out multiplies cost: five round trips, five auth handshakes, five rate limits, and five payloads sitting in context.
• The agent reconstructs correlation across separate MCPs rather than measuring it, because each source crawled at a different time.
• A fused signal like icp_signal (hot, warm, cold) needs all five layers from one observation to be trustworthy.
• Fusion is not a claim to beat every specialist on its own axis. For one deep layer, a specialist MCP can still win.
• Bassethound’s sniff_domain returns tech, infrastructure, firmographics, AI-readiness, and security in a single keyless call.

What does fanning out to five MCPs cost?

Every MCP call is a round trip. Five servers means five resolutions, five TLS handshakes, five tool invocations the host has to schedule. Latency stacks even when the calls run in parallel, because the slowest server sets the floor and each may crawl the target itself. Then there is auth. A keyed infra API, a keyed firmographics API, a keyed tech API, each with its own token, quota, and failure mode. Your agent holds five credentials and handles five ways to fail. Token cost is the quiet tax. Each server returns a full payload, and the host loads all five into context before it can reason. Five verbose JSON blobs can dwarf the question that prompted them. You pay to read them, pay to reconcile them, and pay again every time the agent re-reads context. One call collapses that to a single handshake, one rate limit, and one payload sized to the dossier.

Why can’t the agent just correlate the five answers itself?

It can try. The catch is that the agent reconciles outputs it cannot verify. Server A says the site loads the OpenAI SDK. Server B says the MX record points at Google. Server C reports a security header grade of F. To fuse those into a judgment, the agent has to trust that all three looked at the same site, in the same state, at the same moment. They did not. Each crawled on its own schedule against its own cache. The agent has no way to check whether Server A’s snapshot is a week stale or whether Server B resolved a different edge node. So it correlates on faith and spends reasoning budget doing it. That budget is finite and costly, and the correlation it produces is weaker than one computed at the source, where the crawl provenance is known. Fusion puts the join where the evidence lives. The five detectors read one fetched document set, so they check a link between the AI SDK and the email provider against one observation instead of stitching it across five.

What correlation do you lose across separate crawls?

Timing is the first loss. A domain mid-migration can show one CDN in the morning and another by afternoon. Five crawls at five times give you five truths that never coexisted. Provenance is the second loss. When one crawl feeds every detector, you can say the TLS cert, the DNS TXT vendors, and the AI SDK all came from the same resolution of the same host. Split that across vendors and the shared frame is gone. The third loss is the fused signal itself. icp_signal is not a raw field any single MCP returns. It is a verdict over layers: shipping AI plus a fresh cert plus real firmographics reads hot; an AI headline with no backend fingerprint and an F security grade reads cold. That judgment means something only when the layers describe one company at one instant. A specialist MCP hands you a layer. It cannot hand you the relationship between layers, because it never saw the others.

When is fan-out the right call?

Fusion is not a claim to out-scan every specialist on its own axis. If you need exhaustive port and banner data, Shodan and Censys crawl the whole IPv4 space and hold a corpus no single-domain fetch can match. If you need a deep person-level contact graph, Apollo and ZoomInfo carry identity data Bassethound does not. When your question is one deep layer against a specialist corpus, call that specialist. Fan-out earns its cost when the layers are independent and you want depth in exactly one. The trade flips the moment your question spans layers. “Is this domain a hot AI prospect” is not a tech question or an infra question. It is the correlation. There, five specialist calls give you raw material and a reconciliation chore, and the answer you assemble is weaker than the correlation a single crawl hands you. Pick fusion for the cross-layer judgment. Pick a specialist when you want one axis mined deeper than a general crawl can reach.

How does one call return five correlated layers without five backends?

Because the crawl is the shared substrate. sniff_domain fetches the target once (DNS resolution, HTTP response, headers, TLS handshake, optional JS render) and runs every detector against that one capture. Tech stack, infrastructure, firmographics, AI-readiness, and security are five readings of the same evidence, not five trips to five APIs. Owning the crawl is also why the free tier stays keyless. You do not bring a BuiltWith key or a WhoisXML key, because the server is not proxying them. It does the fetch and the fingerprinting itself, rate-limited. The call is stateless and read-only, so no call depends on the one before it and each dossier is a self-contained observation. Results are cached. Be clear about the limit: a static crawl plus optional render can miss fully server-side or proxied components, and the dossier flags those gaps rather than guessing past them. What you get is one payload, one auth path, one moment in time, and five layers that line up because they came from the same look.

Bassethound perspective

Fan-out is the default because it is easy to build and easy to sell. Every vendor owns one layer, so the ecosystem’s instinct is to wire five servers together and call the pile intelligence. We think that instinct is backwards. The value was never in the five layers. It was in the joins between them, and a join you compute across five independent crawls is a guess with good production values. A competitor who sells a single layer will tell you correlation is the agent’s job. It is not. Handing an agent five payloads and asking it to reconcile snapshots it cannot date is how you burn context and ship a confident wrong answer. Fusion moves the join to the crawl, where provenance is known and timing is one instant. That is a structural edge, not a feature you bolt on, because you cannot correlate layers you fetched separately after the fact. Bassethound returns the relationship, not just the parts: one call, one observation, five layers that describe the same company at the same moment. The dossier is the correlation.

Sources

• Model Context Protocol documentation: https://modelcontextprotocol.io
• MDN Web Docs, HTTP headers: https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers
• Mozilla SSL Configuration Generator: https://ssl-config.mozilla.org
• llms.txt specification: https://llmstxt.org