Every SaaS company is now building its own agent door. Blockworks shipped a hosted MCP server for its datasets. Notolens built one for brand protection. Genea for access control. An insurer launched an entire directory so every carrier can list its own MCP server. Silverchair ran a six-month pilot of one shared endpoint across independent publishers instead of one per publisher.
Stacklok's 2026 report says 45 percent of software organizations already run MCP servers. That is the right instinct. Agents need structured doors, not scraped pages.
But count the cost. Every door is a separate integration to build, secure, version, and maintain. Every API change is an agent breakage waiting to happen. Every new tool in the stack starts the work again from zero. The industry is rebuilding the same connector dozens of times and calling it agent strategy.
One integration per app scales linearly. One infrastructure layer scales once.
If every tool you buy ships its own agent interface next year, who maintains the fifty doors in your stack.
Top comments (1)
This points squarely at the classic O(N x M) integration trap that software engineering rediscovers every decade.
We saw this exact movie with RPC in the 90s (CORBA/DCOM), SOAP in the 2000s, and microservice REST proliferation in the 2010s. The industry starts with high optimism about a standardized protocol format (JSON-RPC over stdio/SSE in MCP's case), but protocol standardization does not solve schema fragmentation. When fifty different SaaS providers each expose their own arbitrary tool definitions, parameter naming conventions, and pagination semantics, the client-side agent still has to shoulder the cognitive burden of navigating fifty bespoke mental models.
Furthermore, there is a hidden security and blast-radius problem with proliferating 'agent doors.' An agent equipped with fifty disjointed MCP servers is effectively operating fifty independent privilege boundaries. Without a unified policy engine or an intermediary orchestration layer that enforces rate-limiting, least-privilege scoping, and transaction rollback semantics, you don't have an 'agent strategy'—you have fifty distributed entry points waiting for prompt injection or cascading tool-call loops.
The future probably won't be fifty bespoke connectors hardwired into every client prompt; it will be unified hub topologies where domain gateways aggregate, normalize, and mediate access, turning a spiderweb of point-to-point connections into a clean, auditable abstraction barrier.
There are only two hard problems in computer science: cache invalidation, naming things, and maintaining 50 different JSON-RPC tool schemas that all implement slightly incompatible definitions of user_id.