New Local AI Gateway Officially Launched: One-Stop Control of Your Models, Usage, and Tools!

Are you still struggling with switching between multiple platform accounts, messy model routing, and painful usage reconciliation? The all-new desktop local AI gateway is here, integrating provider credentials, intelligent model routing, usage history, and local runtime tools into one clean panel, so you can truly control your AI workflow on your own device.

Independent Gateway Panel, Always in Control from Desktop and Tray

A new standalone gateway panel and the system tray's "Usage / Gateway" tabs are now live. You can quickly search and switch your default model with one click, and flexibly add entries to configure various runtime tools. Whether temporarily swapping models or connecting local scripts, the operation is intuitive and efficient.

Custom API providers are supported, along with LAN sharing, freely configurable ports, and one-click model connectivity tests. The model catalog now uses cleaner short aliases while remaining fully compatible with older long IDs, making upgrades seamless.

Security, Flexibility, and Protocols All Upgraded

The credential encryption master key is now stored in a local file at ~/.ai-usage/agent-router/vault.key, so daily use no longer depends on the system keychain; development and release versions can share the same login state, and old keychain entries are migrated once automatically, saving hassle and staying secure.

Gateway request details support local archive viewing (redacted and retained with limits), clearer startup and restart feedback, and further optimized loading skeletons.

Protocol and forwarding capabilities are fully enhanced: native support for mainstream protocols including Messages / Responses / Chat / Gemini, a built-in search execution layer, and full support for client-side tool search and cross-protocol tool invocation. Additional provider capabilities such as Antigravity OAuth, OpenCode Go, and Volcengine cover more real-world scenarios.

DeepSeek Harness can now be connected directly (automatically writing profile and credentials, with support for revocation and restoration); "Runtime Tools" already supports detection and reversible configuration for Claude Code, Codex, ZCode, OpenCode, MiniMax Code, Pi, Claude Desktop, etc., ready to use and easy to roll back.

Multiple forwarding and streaming compatibility issues for Claude / Codex / WorkBuddy / HY4 have also been fixed; when the old version of the gateway returns a plain-text 404, the system now gives a clear, actionable restart prompt to reduce troubleshooting time.

Usage and Subscriptions, More Precise and Transparent

Usage collection has been upgraded again: new CodeArts Agent and PI-Desktop usage collectors added; Trae international accounts can read quotas and usage directly from an existing IDE login, without requiring an additional session library key.

A dedicated Grok Bot subscription quota card has been launched (no Cursor installation needed), and Qoder quota pools now uniformly display remaining percentages, giving you a clear view of remaining resources.

Opening the app no longer proactively pops up the Cursor keychain authorization prompt: after the first refresh and authorization on the usage card, quotas will display automatically later, making the experience quieter.

Sync feedback has been further refined, clearly distinguishing between local collection complete, cloud upload complete, pending report, failure, and unreported states, avoiding showing "success" when the cloud hasn't updated yet.

The tray's "Usage Overview" supports showing/hiding metrics individually and remembers the local selection; it also fixes issues like continuous loading after quick time-range switching, and interaction problems with heatmaps and filter bars, making daily viewing smoother.

Latest official channel pricing from models.dev is synchronized (covering Claude Fable, Gemini 4 Carbon, Muse Ember, GPT-6.1 Terra, Mistral Large 4, etc.) for more accurate cost estimation.

Other Optimizations

• Fixed DSH Desktop custom data directory and Beta sessions not being discoverable for statistics;
• Limited the number of incremental deduplication and snapshot retention for Codex / ZCode to prevent cursors.json from growing continuously;
• Fixed an issue where macOS installer retry builds mistakenly collected intermediate disk images.

From model routing to usage reconciliation, from local tool integration to secure credential management, this update turns "usable" into "good to use." Whether you're a heavy multi-model user or a team needing fine-grained usage control, the new local AI gateway makes your workflow lighter, more stable, and more controllable.

Update now and put AI truly in your own hands.

Download link https://github.com/juejin-cn/juejin-usage/releases/tag/v1.1.0