What's new in AI-Proxy

The "everything just works" release — every fix below was verified by execution on real macOS, Windows, and Linux hardware before this build was allowed to ship.

Fixed

• Provider sign-in could hang forever after a cancelled attempt. The OAuth callback is now serialized per provider, a busy callback port shows an honest error instead of an endless spinner, and cancelling truly cancels — on every platform.
• Removing a provider login now actually removes it. The engine forgets the token immediately (no more logins resurrecting themselves after you deleted them).
• macOS: the remove and Reset buttons did nothing. Confirmation dialogs were silently swallowed by the system webview; both flows now use an inline two-step confirm — first click arms, second click executes.
• Windows: the uninstaller's "Delete application data" checkbox now wipes the real data folder. Reinstalls no longer come back signed-in.
• The management API no longer locks itself out after your first provider login.

New

• Everything is free — just sign in. One-click Configure and Copy values for every client now require only a (free) account, not a paid plan.
• Buzz support (buzz.xyz — Block's team-chat-with-agents workspace): a one-click Configure points Buzz's OpenAI-compatible provider at your gateway; pick a model inside Buzz and your agents run on your subscriptions.
• ARM64 Linux, for the first time: AppImage + .deb + headless CLI for Raspberry Pi, Graviton/Ampere boxes, and Apple-Silicon VMs.
• Kind-aware Models Catalog: each tab suggests only what's real — text models (with tool-calling badges), vision models, and honest placeholders for the upcoming local speech backend.
• Reset app data on the Server screen — factory-fresh in two clicks.

Auto-update starts here

0.3.2 installs will offer this update in-app — one click and you're on 0.3.3. From now on, every release arrives by itself.

Under the hood

• Every artifact is scanned before publish — no state files, no secrets, no PII can ship.
• A provider-free functional regression (15 live checks) now runs in CI on every build.
• Local speech-to-text and text-to-speech (Whisper/Piper-class, free and offline) is designed and spiked — coming in an upcoming release, powering dictation workflows through the standard /v1/audio API.

Cross-platform AI-Proxy: macOS dmg, Windows msi, Linux AppImage/deb + headless CLI tarballs.

Cross-platform AI-Proxy: macOS dmg, Windows msi, Linux AppImage/deb + headless CLI tarballs.

AI-Proxy Desktop 0.3.0 — JSON mode, per-key usage, the on-prem release

Completes the P2 surface on top of 0.2.9's DVT-first-class gateway:

• response_format: {"type":"json_object"} now actually works — providers ignore it upstream, so the gateway enforces it: instructs the model, validates, strips fences, retries once. Tool-call turns are never touched (agent loops stay intact — adversarially verified).
• Per-key usage attribution — GET /v1/usage (any valid key) returns per-key request + token totals, keyed by fingerprint (…last6; full keys never stored). Add one key per user/team in api-keys:; optional names via GATEWAY_KEY_NAMES. Counters persist across restarts (atomic writes).
• On-prem controls — gateway bind address honors host: in config.yaml; guide/08 gains the floor playbook: per-user keys, /health monitoring, launchd/systemd supervision, Caddy/nginx TLS fronting (with the SSE proxy_buffering off gotcha documented).
• Verified: real usage totals in responses; OpenAI SSE streaming with [DONE]; regression suite now 15 checks — all green.

Install: brew install --cask meta-thinking/tap/ai-proxy · Docs: https://aiproxy.meta-thinking.net/docs macOS Apple Silicon. (Intel / Windows / Linux ship via CI.)

AI-Proxy Desktop 0.2.9 — the first-class DVT gateway (plus the Clients revamp)

Gateway hardening (every finding from the DVT integration, fixed + regression-locked):

• System role honored everywhere — Claude-routed requests fold system content into the first user turn (their harness used to silently discard it); Gemini untouched. Your agent's grounding always reaches the model.
• Truthful /v1/models — retired ids pruned/remapped (2 newly-dead Anthropic ids added); an explicit unknown model now fails fast with OpenAI-shaped 404 model_not_found; an omitted model always completes on a live default.
• Clean error contract — every error leaves in OpenAI shape with correct status + actionable message (auth_unavailable → 503 + Retry-After). Sampling params a model rejects (temperature on newest Claude) are dropped & retried by the gateway.
• GET /health — token-free, key-free: {status, default_model, providers:{claude|gemini|openai|ollama: ok|cooldown|no_auth}}.
• Service supervision — ai-proxy run (foreground), ai-proxy service-install (launchd auto-start + auto-restart).
• Regression suite: scripts/regress.sh — 9/9 (system role, model sweep, param-drop, 404, tool round-trip, concurrent isolation, /health).

Connect Clients revamp: 38 clients with logos in 6 categories, installed-client detection, 12 battle-tested config writers with backup/Undo — DVT featured first-class (writes ~/.dvt/ai_keys.yml, byte-identical undo).

Install: brew install --cask meta-thinking/tap/ai-proxy · Docs: https://aiproxy.meta-thinking.net/docs macOS Apple Silicon. (Intel / Windows / Linux ship via CI.)

AI-Proxy Desktop 0.2.8 — Tool calling / agents

The OpenAI function-calling protocol now passes through the gateway end-to-end: tools[], tool_calls, finish_reason:"tool_calls", streamed tool-call deltas, the role:"tool" round-trip, tool_choice, and parallel calls — all on your own subscriptions. Point opencode or the n8n Tools Agent at the proxy and your agents take real actions.

• /v1/models now tags each model "tools": true|false so clients can filter to tool-capable models.
• The retired-model remap rewrites only the model field (byte-safe), so your tools/messages arrays are never re-encoded or mangled.
• Best for agents: claude-sonnet-4-6 / claude-opus-4-8 (forced tool_choice + parallel). gemini-3-flash for single tool calls.

Install: brew install --cask meta-thinking/tap/ai-proxy · Docs: https://aiproxy.meta-thinking.net/docs macOS Apple Silicon. (Intel / Windows / Linux ship via CI.)

Dead/no-model candidate-cascade (fixes DVT + any client relying on a default model). macOS arm64.

Client Configure/Undo (real file writers, backed up), runtime dead-model self-heal, custom providers, launch-at-login, richer tray. macOS arm64.

Parity restore: custom providers, launch-at-login, reveal config/log, richer tray, meaningful refresh + post-login model rescan, retired-model fixes. macOS arm64.

New standalone account backend, Free-forever + Pro model, brand icon, custom port, retired-model catalog fixes. macOS arm64 (Intel/Windows/Linux return with the CI pipeline).

Cross-platform AI-Proxy: macOS dmg, Windows msi, Linux AppImage/deb + headless CLI tarballs.

Cross-platform AI-Proxy: macOS dmg, Windows msi, Linux AppImage/deb + headless CLI tarballs.