MCP spec update lands as benchmark shows models still depend heavily on docs
craigsdennis · x · 2026-07-29
Cloudflare is being pitched as a strong home for MCP servers as the new MCP spec lands with stateless-by-default behavior, multi-round-trip elicitations, and CIMD auth by default.
The post links to mcpbench, a benchmark for how well models can act as MCP clients and servers. The results break performance down by capability and show a stark gap between full docs and no docs:
- Client performance improves dramatically with documentation, especially around authorization, lifecycle/session handling, and async tasks.
- Server performance is also much higher with docs, with top models reaching near-ceiling on some capability buckets.
- The benchmark ran 38 client scenarios / 516 trials and 16 server scenarios / 228 trials.
- The author’s point is that the spec changes are real, but models still depend heavily on docs to behave well as MCP participants.
More from coding & agent
- Samsung uses Claude Code for chip verification: one-month tasks done in two days, 15x speedup — 新智元 · 2026-08-25
- Swarms MCP Launches: Single Endpoint for Multi-Agent Workflows — KyeGomezB · 2026-08-25
- Vetta Harness Cuts Long-Horizon Agent Costs by 66% vs. Claude Code — ycombinator · 2026-08-25
- How to set up Agent evaluation: scripted checks or expert review? — mastra_ai · 2026-08-25
- Dev Seeks Less Invasive AI Models for Spelling and Grammar Checks — mitsuhiko · 2026-08-25
- Google Antigravity generated a scenario-aware Flutter UI for a web app in minutes — rseroter · 2026-08-25