Skip to content
AI Primer
release

OpenClaw 2026.4.27 adds DeepInfra support and forward-proxy routing

OpenClaw 2026.4.27 bundles DeepInfra support, better non-image attachments, explicit forward-proxy routing, and stricter model selection. The update broadens provider access while hardening operator-run deployments against routing and session failures.

4 min read
OpenClaw 2026.4.27 adds DeepInfra support and forward-proxy routing
OpenClaw 2026.4.27 adds DeepInfra support and forward-proxy routing

TL;DR

  • openclaw's release thread says OpenClaw 2026.4.27 bundles DeepInfra, adds forward-proxy routing, tightens model selection, and ships a long list of reliability fixes.
  • According to openclaw's DeepInfra post and the Model providers docs, DeepInfra now sits in the bundled provider set with OpenAI-compatible model discovery plus image generation, editing, media understanding, TTS, text-to-video, and embeddings.
  • openclaw's attachment update pairs with the WebChat docs: chat.send can now stage non-image attachments for agents, and the WebChat or Control UI upload path no longer trashes chat state on non-video files.
  • As openclaw's proxy note describes it, operators can force runtime HTTP and WebSocket traffic through an explicit forward proxy, while openclaw's model-selection note says unreachable user-picked models now fail visibly instead of silently falling through.

You can read the full v2026.4.27 release notes, browse the updated model selection docs, and trace the quieter platform work in openclaw's Matrix approval post and openclaw's reliability follow-up. The fun headline is DeepInfra. The more operator-brained changes are the proxy knob, the exact-model failure behavior, and the steady removal of silent weirdness.

DeepInfra

OpenClaw's new bundled DeepInfra integration is broad enough to matter even if you never touch the provider manually. openclaw's post lists dynamic OpenAI-compatible model discovery, image generation and editing, media understanding, TTS, text-to-video, and embeddings behind one key.

The official release notes add one under-the-hood detail the tweet only hints at: DeepInfra ships with provider-owned onboarding policy, part of a broader move toward manifest-first provider metadata. The Model providers docs already show deepinfra as a first-class provider ID with DEEPINFRA_API_KEY auth and an example deepinfra/deepseek-ai/DeepSeek-V3.2 model path.

Attachments and model selection

The file-handling change is small but real. openclaw's attachment update says chat.send now stages non-image attachments as agent-readable media paths, while the release notes say unsupported RPC attachment paths are now handled explicitly instead of silently dropping files.

That lands next to a second anti-surprise fix. openclaw's model-selection note says explicit /model choices and picker selections now error visibly when the target is unreachable. The Models CLI docs use almost the same language: user session selections are exact, and if the selected provider or model fails before producing a reply, OpenClaw reports the failure instead of answering from an unrelated fallback.

For local models, openclaw's note also calls out leaner Ollama probes and preserved thinking metadata. The Ollama docs say OpenClaw reads /api/show capability data when it can, including thinking, and falls back to model-name heuristics when Ollama omits those fields.

Proxy routing

Forward-proxy support is the most obviously operator-facing addition in this release. openclaw's proxy note says runtime HTTP and WebSocket traffic can now be routed through an explicit outbound proxy, with strict http:// validation and a loopback Gateway bypass.

The release notes frame that as opt-in, operator-managed outbound proxy routing, plus cleanup of proxy state on exit. That combination reads like a deployment hardening pass more than a convenience feature, especially for teams running OpenClaw behind controlled egress.

Reliability and presence

OpenClaw tucked a lot of practical cleanup into the same ship. openclaw's reliability follow-up groups together faster gateway and channel startup, safer redaction, better session cleanup, tighter update behavior, and fewer retry loops.

The release notes add two details worth surfacing separately:

  • plugin startup and model catalogs are moving toward manifest-first metadata, which reduces Gateway boot work and makes provider rows, aliases, and suppressions easier to audit
  • iOS and Android now publish authenticated node.presence.alive events plus last-seen metadata, so background wakes can mark a node recently alive without treating it as fully connected

openclaw's Matrix approval post rounds out the same theme on the client side: richer approval metadata for Matrix clients, plus tool-progress updates in live preview edits when preview streaming is enabled.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 3 threads
TL;DR1 post
DeepInfra1 post
Reliability and presence1 post
Share on X