algis-hn

Accurate for naive MCP client implementations, but a proxy layer with inference-time routing solves exactly this control problem. BM25 semantic matching on each incoming query exposes only 3-5 relevant tool schemas to the agent rather than loading everything upfront - the 44K token cold-start cost that the article cites mostly disappears because the routing layer is doing selection work. MCPProxy (https://github.com/smart-mcp-proxy/mcpproxy-go) implements this pattern: structured schemas stay for validation and security quarantine, but the agent only sees what's relevant per query rather than the full catalog. The tradeoff isn't MCP vs CLI - it's routing-aware MCP vs naive MCP, and the former competes with CLI on token efficiency while retaining the organizational benefits the article argues for.

Hard caps treat the symptom. If agents call tools they shouldn't need, the root cause is tool set visibility - 80 tools in context means the model tries irrelevant ones, fails, and retries. Pre-filtering to a semantically relevant subset per request (BM25 on the incoming query works well) eliminates most retry loops before they start; the cap stays as backstop for genuinely pathological cases.

The auth problem you've hit is the exact use case for OAuth relay at the proxy layer - the proxy holds enterprise tokens and handles refresh independently, so each tool integration stays stateless and free of credential lifecycle complexity. The natural language to specific ID mapping (your channel_id example) is the genuinely harder piece, though semantic routing at the middleware layer helps: the proxy selects the right tool based on semantic matching, which reduces how often the LLM needs to reason about entity IDs directly. What's missing isn't MCP features - it's an opinionated middleware layer that decouples auth handling and semantic routing from individual tool implementations. Some open-source proxy projects are building exactly this, though enterprise connectors for Teams and Salesforce at scale are still thin.

110 tools is exactly where you need a discovery layer in front. A proxy that only surfaces the 5-10 relevant tools per query keeps the context clean without splitting into separate servers https://github.com/smart-mcp-proxy/mcpproxy-go