Agents get a voice, a CLI console, and a seat at the table

Voice enters the AI Gateway

The biggest story of the week is Vercel's expansion of the AI Gateway into voice and audio. Realtime voice, speech-to-text, and text-to-speech are now available through the AI SDK, with the same observability, routing, and spend controls as text and image models. The launch includes xAI's Grok audio models — Grok Voice Think Fast for realtime, plus dedicated TTS and STT models — all accessible via the AI SDK 7. This is a significant capability unlock: developers can now build voice agents with the useRealtime hook and manage them through the same gateway infrastructure they already use for chat and completion workflows.

Across the AI SDK, the week also brought a steady stream of polish. The OpenAI provider gained support for the web_search_call.results include option, the MCP client added Streamable HTTP session hooks with reattachment support, and the Gateway provider introduced a has option to restrict routing by model capabilities. A potential path traversal vulnerability in the Pi harness adapter was fixed, and two provider-utils patches addressed unbounded memory use on JSON response body reads.

The CLI becomes a data console

Vercel's vercel metrics command, which shipped earlier on the Speed Insights side, this week also pulled Web Analytics data into the same interface. The command now returns page views, visitors, custom events, and core Web Vitals (LCP, INP, CLS, FCP, TTFB) — all queryable from the terminal. For teams that rely on agents or CI pipelines to answer traffic questions, this eliminates the need to open a dashboard for basic performance queries.

The CLI also saw meaningful infrastructure work. Service auto-detection migrated to V2 format, with layout detection now resolving via the V2 resolver and generating per-service path transform routes. The Python runtime added WebSocket support for WSGI apps, exposing the raw connection socket for upgrade requests in Flask and similar frameworks. And the Vercel Connect library began auto-provisioning connectors on first use inside deployments, reducing the OAuth setup friction for managed integrations.

Cloudflare: agents, jurisdiction, and cache control

Cloudflare's Agents SDK v0.17.0 introduced detached background sub-agents that run durably with live progress reporting, without blocking the calling turn. A new runTurn() facade unifies turn admission behind a single mode parameter, and a wave of reliability fixes continues converging the SDK's chat and agent models. For teams building long-running agent workflows, the background sub-agent pattern is a meaningful step toward production-grade autonomy.

On the data residency front, Durable Objects now support a us jurisdiction that restricts compute and storage to the United States, with Workers able to access them from anywhere. Separately, Workers fetch() requests can now set cf.vary to control how Cloudflare caches origin responses with a Vary header — a narrow but useful knob for cache tuning at the subrequest level.

Finally, Cloudflare One added service token support for MCP server portals, allowing autonomous agents to connect without a browser-based OAuth flow. It's a small change that points to a larger trend: as agent-to-server communication becomes routine, platforms are quietly building the auth infrastructure to support it.