The Pattern
Anthropic's Model Context Protocol, released in late 2024, promised to standardize how LLMs connect to enterprise data and tools. Block, Apollo, Zed, Replit, and Sourcegraph adopted it. But a counter-narrative is emerging from teams who've tried to implement it: MCP solves the wrong problem.
The core complaint: to expose an existing, stable business API to an AI agent, MCP requires building a compliant server layer, wrestling with JSON-RPC 2.0 transport protocols, and mapping resources. "This isn't integration; it's building a massive wall around your business system," argues one implementation critique gaining traction in developer communities.
The Real Costs
Three issues surface in production:
Dual maintenance burden. Teams maintain both original business logic and separate protocol translation logic. Double the documentation, double the test cases, double the bug surface area.
Context consumption. Tool definitions bloat prompts with metadata. When an LLM parses an MCP server's tool list, verbose protocol text consumes the context window. Result: higher token costs, slower responses. In high-concurrency scenarios, this matters.
Debugging complexity. When a call fails, is it prompt drift? Protocol misconfiguration? Underlying business logic? Each architectural layer doubles troubleshooting entropy.
The Alternative
The "Skill Pattern" - championed by frameworks like Solon AI - takes a different approach. Instead of adapting business logic to a protocol, it wraps existing capabilities:
- Annotated methods become callable skills
- Existing Swagger/OpenAPI docs import directly
- Database connections with read-only permissions enable text-to-SQL
One example: a REST API skill that reads Swagger documentation from a URL, no new code required. Another: database access that lets AI generate and execute SQL queries within security constraints.
What This Means
MCP adoption resembles past architectural rushes - the microservices wave that saw teams fragment working monoliths for pattern-matching rather than solving actual problems. History suggests: evaluate whether the protocol solves your integration challenge, or whether simpler approaches suffice.
The protocol's benefits - standardized AI-to-system connections, reduced custom integration work - are real. But so is the implementation tax. For teams with stable APIs and clear business logic, lighter-weight approaches might ship faster.
Worth noting: no major MCP adopter has published implementation cost analyses or post-deployment metrics. The pattern is young. We'll see which approach scales when enterprise deployments hit production load.