Skip to content

AI-Proxy 0.3.4

Choose a tag to compare

@heshamh96 heshamh96 released this 23 Aug 13:35
· 2 commits to main since this release

The release that makes AI-Proxy honest, multi-modal, and truly cross-platform. Every fix below was verified by execution on real macOS, Windows, and Linux hardware before shipping.

The catalog now tells the truth

  • Only models that actually work are listed. AI-Proxy now probes each provider and lists a model only once it has proven it can serve — dead subscriptions (out of credits, no membership) vanish from /v1/models instead of failing mid-request. /health shows an honest per-provider verdict (ok / cooldown / blocked:quota).
  • No more phantom image/video models from providers that can't actually serve them.

Images work in your coding agents

  • Image models (e.g. gemini-3.1-flash-image) now return a rendered image your agents can display — including over streaming, which is how every agent (Buzz, opencode, Goose) talks to the gateway.

Local speech — free, offline, forever

  • Speech-to-text (Whisper) and text-to-speech (Piper/Kokoro) ship built in. Download a model from the Catalog's Audio tab and use it through the standard /v1/audio/* API — no account, no cloud, no cost. Foundation for dictation and voice workflows.

Connect Clients, rebuilt

  • New layout: toggle between By status (Installed on top after a scan) and By type, with search — the Models-Catalog experience, for your tools.
  • Buzz is a first-class client on all three OSes now (the Windows/Linux config paths are fixed).
  • One-click Configure is correct cross-platform at last: opencode, Goose, and llm now write to the right place on Windows (%APPDATA%), and the shell key-setup works in both zsh and bash on macOS and Linux — previously it silently missed bash/Linux users.
  • "Re-point all": after a key change or reset, one click re-points every configured tool to the current key. A factory reset no longer forgets which tools you'd set up.

Everything is free — just sign in

  • All one-click Configure and Copy-values actions now need only a (free) account, not a paid plan.
  • Sign in with Google is now supported alongside email.

Fixes

  • The provider sign-in hang is gone (per-provider OAuth serialization + honest errors instead of an endless spinner).
  • Removing a provider login now truly removes it.
  • macOS: the Remove and Reset buttons work (they used an inline two-step confirm now).
  • Windows: the uninstaller's "Delete application data" wipes the real folder; a rare "Start does nothing" state self-heals.
  • The Windows auto-updater no longer stalls on a hidden dialog (sidecars are stopped before the update applies).

Auto-update

  • 0.3.3 installs will offer this update in-app — one click to 0.3.4.

Under the hood

  • Every artifact is scanned before publish (no state, secrets, or PII can ship).
  • A provider-free functional regression + a "listed ⊆ servable" soak test run in CI on every build, so a dead-model or broken-catalog regression can never reach you again.