OpenClaw ships browser realtime transport, Cerebras, asymmetric embeddings (2026.4.26)

What changed

OpenClaw 2026.4.26 (April 28, 2026) adds a generic browser realtime transport contract along with Google Live browser Talk sessions and a Gateway relay for backend-only realtime voice plugins. Cerebras joins as a bundled plugin with onboarding, static model catalog, docs, and manifest-owned endpoint metadata. Memory and OpenAI-compatible plugins gain optional memorySearch.inputType, queryInputType, and documentInputType config for asymmetric embedding endpoints, including direct query embeddings and provider batch indexing. Pre-runtime model-id normalization, provider endpoint host metadata, and OpenAI-compatible request-family hints move into plugin manifests so core no longer carries bundled-provider routing tables.

Why it matters

The asymmetric embedding config is the production-facing win. Retrieval pipelines using Cohere v3, Voyage, OpenAI text-embedding-3, or NVIDIA NV-Embed all benefit from different embedding modes for queries vs documents; previously you had to fork the OpenClaw plugin or hand-roll the split. Native config means cleaner upgrades and consistent retrieval quality. The browser realtime transport contract sets up a clean abstraction for the wave of realtime voice provider APIs landing this year.

What to watch for

If you run RAG over Cohere or Voyage embeddings, audit your pipeline against the new queryInputType config and re-benchmark retrieval. If you build voice products on OpenClaw, prototype against the new browser realtime transport so you stay portable across providers. Plugin authors should track the manifest-first migration and move any hardcoded routing logic out of core overrides into manifest fields.