Quandri Engineering: Why Skills and CLIs Beat MCP for Most Dev Workflows

Quandri’s backend team measured the real cost of Model Context Protocol servers against their stack and found the architecture struggles to justify itself for typical developer work. Four connected MCP servers consumed 10.5% of the context window in tool definitions alone, with Linear’s 42 always-loaded tools eating over 12,800 tokens before a single call. Looking up one Linear issue cost roughly 65x more tokens through MCP than a direct curl against the GraphQL API, and benchmarks from the original referenced post showed MCP calls running 3x slower than equivalent REST hits due to the extra process layer.

The team’s preferred pattern is to lean on existing CLIs (gh, psql, aws) for daily tools and wrap them in Skills that load only when invoked, keeping tool instructions off the desk until needed. MCP still earns its place for services with no real CLI, non-developer users, real-time bidirectional flows, and database access where server-side guardrails can block destructive queries or hide credentials. The author notes Claude Code’s new Tool Search with Deferred Loading reportedly cuts schema bloat by 85%+, which addresses the context complaint but not the latency, debugging friction, or the marketing-driven proliferation of half-baked MCP servers.

The broader argument is that the industry is repeating the ‘AI-powered’ and ‘blockchain-based’ checkbox cycle, with vendors shipping MCP support regardless of stability. Quandri’s switch to Skills-plus-CLI freed about 21K tokens and removed initialization failures from their workflow, suggesting that teaching the model to use tools it already understands beats bolting on a new protocol layer.