Why can't my agent just call five MCPs?

Best move: call one tool that returns five correlated layers, not five tools that each return one.

Why it works: the value is the join across layers, and a join only holds when every layer comes from the same crawl. Five separate MCPs each crawl at a different moment, so their outputs never line up.

Key takeaways

• Five MCPs mean five round trips, five auth flows, five rate limits, and five response schemas your agent has to reconcile.
• Correlation across layers only holds inside one crawl snapshot; independent tools crawl at different times and produce views that do not line up.
• One call returns a fused summary (primary_stack, ai_readiness verdict, security_grade, icp_signal), so the agent gets a decision instead of raw material.
• Deep single-layer forensics (port scans, full whois history) is still a job for a specialist tool; fusion is for qualification, not forensic depth.
• Bassethound runs the investigation inside one request, keyless, and reports the layers it could not reach rather than guessing.

What breaks when you fan out to five tools?

Five MCPs mean five connections, five tool schemas loaded into your agent’s context, and five response formats to parse. Each server has its own auth, its own rate limit, and its own idea of what a domain lookup returns. Your agent now owns an orchestration problem: plan the order, fire the calls, wait on the slowest one, and handle the case where two succeed and three time out. Latency is additive unless you parallelize, and parallelizing adds its own failure handling. Then comes reconciliation. Tool A calls it hostname, tool B calls it host, tool C nests it three levels deep. You write glue code, or you burn tokens asking the model to normalize on the fly. None of this work produces signal. It is tax you pay to get five disconnected views into one prompt.

And you still have not joined them. You have five payloads sitting in context, each describing the domain from a different angle at a different moment, waiting for someone to make them agree. That someone is your agent, doing expensive reasoning over data nobody built to line up.

Why can’t the agent correlate the five results itself?

Because correlation needs a shared snapshot, and five tools do not share one. Each MCP crawls the domain on its own schedule. One follows redirects, one does not. One renders JavaScript, one reads raw HTML. One returns a cached result from an hour ago, one hits the site live. So when tool A reports a TLS certificate and tool B reports a tech stack, they may be describing two different states of the same site. Join them and you get a confident answer built on mismatched inputs. The join is the product. Reading that a domain loads the Anthropic SDK is a fact. Reading that the same crawl also found a Pinecone client, a live MCP endpoint, and a valid llms.txt is a verdict: this company ships AI. You only earn the verdict when every layer comes from one fetch of one page at one time. An agent stitching five outputs cannot manufacture that guarantee after the fact. It can guess, and the guess will read as certainty, which is worse than an honest gap. Fusion happens at crawl time, inside one request, or it does not happen.

Doesn’t calling five tools cost more tokens and latency?

Yes, on both. Every tool you expose to the model adds its schema to the context window before the conversation starts. Five domain tools mean five schemas the model reads on every turn, whether it calls them or not. Then each call returns a full payload, and a domain dossier is not small. Five payloads land in context, and the model holds all of them while it reasons. Now add the reconciliation prompt, because the shapes do not match. You are paying for tokens that carry structure, not signal. Latency stacks the same way. A sequential fan-out waits on five servers in a row, gated by the slowest and the flakiest. A parallel fan-out cuts wall-clock time but hands you five concurrent failure modes to handle, plus retry logic. One call collapses all of it: one schema, one payload, one round trip, one scored summary the model can act on. The five-layer dossier is denser than any single response, but it replaces five responses, and it arrives pre-joined so the model spends its tokens deciding, not stitching.

When is fan-out the right call?

Fusion is not always the answer, and pretending otherwise would be dishonest. If you need forensic depth in a single layer, call the specialist. Shodan and Censys index the whole internet’s port and service history at a depth no general dossier matches. DomainTools carries whois pivot history and passive DNS going back years. Apollo and ZoomInfo hold contact graphs no crawl can reconstruct. Ask for every host that ever served a cert, or the full buying committee, and you have a corpus job. The corpus lives in one specialized tool. Fan out to it. The distinction is qualification versus investigation. Bassethound answers “is this domain worth my agent’s time, and what does its stack look like right now” in one keyless call. It qualifies. When a lead clears that bar and you need to go deep on one axis, reach for the specialist that owns that axis. The mistake is not using five tools. The mistake is using five general tools to reconstruct a correlation you could have gotten pre-joined from one, and calling the reconstruction a shortcut.

Bassethound perspective

The market assumes composition is free. Buy a tech-stack tool, an infra tool, and a firmographics tool, wire them into your agent, and let the model glue the outputs. Every vendor in this space prices and shapes itself around that assumption, because each of them owns one layer or one corpus and wants to be one of your five. We dispute the assumption. Composition is not free. It is the most expensive part, and you are the one paying it, in tokens, in latency, and in silent correlation errors your agent will state with total confidence. The join across layers is where domain intelligence lives, and a join only holds inside a single crawl. You cannot buy that from five separate tools, no matter how good each one is. So Bassethound does the fan-out once, server-side, and hands your agent the correlated result: five layers, one snapshot, one scored verdict, keyless. We are honest about the edge. A static crawl with optional JS render can miss fully server-side or proxied components, and the dossier says so instead of guessing. Fusion is the moat. Fan-out is the tax.

Sources

• Model Context Protocol documentation: https://modelcontextprotocol.io
• Model Context Protocol specification: https://modelcontextprotocol.io/specification
• MCP tools concept: https://modelcontextprotocol.io/docs/concepts/tools
• HTTP overview (MDN): https://developer.mozilla.org/en-US/docs/Web/HTTP/Overview