MCPCon lays out MCP roadmap and OpenAI best practices for servers

At the AGNTCon/MCPCon keynote, the official MCP roadmap was presented alongside OpenAI's best practices derived from ChatGPT production data, signaling that MCP has moved from protocol specification into large-scale real-world traffic and ecosystem governance.

Confirmed

• The roadmap names five priority areas: Agentic messaging (MCP Tasks for long-running async tasks + MCP Events for server-initiated pushes/webhooks), richer semantics (Skills over MCP), and more, each at different maturity levels.
• MCP Tasks allows tasks to run from seconds to days; clients can disconnect and check status later, currently via polling. MCP Events, in development, will provide a push channel.
• Elicitation is already available in Claude: MCP servers can ask the user questions mid-execution, e.g., to confirm destructive operations, fill in missing configuration, or redirect to login/authorization.
• Skills over MCP, in development: servers deliver domain knowledge alongside tools, with skills tied to the logged-in identity — different roles (finance, engineering, design) hitting the same backend get different workflows. Pamela Fox relayed dsp's point that authenticated MCP servers can personalize skills per logged-in user or even per individual.
• File uploads: 150+ MCP servers implement custom upload flows and 100+ offer download tools, but the flows are mutually incompatible; a working group is pushing for a unified standard.
• OpenAI best practices: only about 56% of MCP servers provide server instructions — keep them within 1000 characters with key info up front to avoid ChatGPT truncation; about 61% of tools declare annotations, and verified read-only tools can reduce security screening and speed up invocation; expose each tool as a clear, individual action rather than hiding them behind dispatcher patterns like listCommand/executeCommand (the extra dispatch step adds latency); about 3/4 of tool schemas are under 1KB — oversized schemas may be truncated, so check the actually generated schema before shipping.
• Production traffic: Resend reported 1 million MCP calls in a single month, its fastest-growing usage category; Railway's MCP daily users grew over 300% from June to July, then another 500% from July to August.

Why it matters

• OpenAI's data comes from ChatGPT's real production environment, so its recommendations (instructions, annotations, schema size) can directly guide MCP server authors in improving invocation rates and latency.
• Tasks/Events/Elicitation/Skills together fill gaps in long-running agent tasks, human-in-the-loop confirmation, and domain knowledge injection, positioning MCP servers as a distribution channel for skills.
• Fragmented implementations like file uploads show the ecosystem urgently needs standardization, and the roadmap points a clear way forward.

2026-10-23 ~ 2026-10-23 · 12 related posts

• Auth'd MCP servers can deliver user-personalized skills, an overlooked design win — dsp_ · 2026-10-23
• [source] MCP roadmap: five priority areas from agentic messaging to SDK DX — bibryam · 2026-10-23
• MCP Tasks runs jobs for hours or days; Events adds server push — bibryam · 2026-10-23
• Skills over MCP in development: domain knowledge tied to your login — bibryam · 2026-10-23
• MCP Elicitation lands in Claude: servers can ask users questions mid-action — bibryam · 2026-10-23
• [source] Resend logs 1M monthly MCP calls as OpenAI frames MCP as a distribution channel — bibryam · 2026-10-23
• ~76% of MCP Tool Schemas Are Under 1 kB, but Oversized Ones Risk Truncation — bibryam · 2026-10-23
• OpenAI: Don't Hide MCP Tools Behind listCommand/executeCommand Dispatchers — bibryam · 2026-10-23
• OpenAI: Honest Tool Annotations Mean Less Screening and Faster MCP Calls — bibryam · 2026-10-23
• [source] Only ~56% of MCP Servers Provide Instructions, OpenAI Data Shows — bibryam · 2026-10-23
• OpenAI: Keep MCP Server Instructions Under ~1,000 Characters or ChatGPT May Truncate Them — bibryam · 2026-10-23
• MCP file uploads: 150+ servers with custom flows, working group seeks a standard — bibryam · 2026-10-23