MCP's new roadmap: progressive tool discovery, agent identity, tasks
The core maintainers just set five priorities — from HTTP transport to killing API keys — that reshape MCP servers. Plus: a Claude Code hive-mind on HN.
Copy markdown
Progressive tool discovery kills the tool-catalog dump
Servers will expose a small entry point first instead of flooding the model with every tool upfront — the roadmap's fix for the context bloat and mis-selection that hits any MCP server with a big catalog.
Agent identity replaces API keys
Servers will recognize which agent is calling via existing standards like DPoP and Workload Identity Federation, not shared keys. For anyone shipping MCP to a team, this is the path to per-agent auth and audit you don't have to hand-roll.
One transport: HTTP-native, stdio included
Building on the 2026-07-28 spec, the roadmap unifies remote HTTP and local stdio servers under a single transport — less connection glue to write in every client and server.
Agent loops become first-class
The Tasks extension matures and server-initiated events get standardized, so MCP moves past request/response to long-running work a server can push updates on and you can steer mid-flight. SEPs aligned to these five priorities get expedited review.
Elsewhere: Munder Difflin wires Claude Code terminals into a hive mind
A trending open-source (MIT) desktop harness wraps real Claude Code terminals as agents with shared memory and mailboxes under a "GOD" orchestrator agent. It hit the HN front page today — macOS, Node 18+, and Claude Code on your PATH to try it.