Skip to main content

​
Changelog

Stay up to date with the latest changes to Redbit AI Studio.

​
Unreleased

​
🔄 Model refresh (2026-09)

  • Catalog verified against live relays — every generator slug was re-checked against each relay’s own published catalog: AtlasCloud and WaveSpeed live APIs, AceData’s OpenAPI model enums, and Grsai’s API docs + public catalog (pnpm audit:models, 142 slugs). New generations added: Seedance 2.5, SeeDream V5 Pro, Wan 3.0 / 3.0 Prime, GPT Image 2.5 (Sunburst + Flare), Kling V3.0 Motion (Std/Pro), Midjourney V8.2, MiniMax Speech 2.8, Suno Chirp V6.
  • Relay-specific model fixes — AceData was being sent Seedream/Seedance ids it does not publish (doubao-seedream-5-0-260128) and a gemini-3.0-pro describe id that never existed; Grsai was sent retired gpt-image-1* ids and a gemini-3.1-pro-preview id it names gemini-3.1-pro. All of them now resolve to ids the relay actually serves.
  • Picking a model now always calls that model — SeeDream V4.5 on AtlasCloud and Seedance 2.0 Mini on WaveSpeed had no mapping row, so the request silently fell back to V4 / standard 2.0 while the UI still showed (and billed) the selected variant. A new completeness test fails the build whenever a shipped variant has no mapping for a relay it can be sent to.
  • Retired upstream models removed — Suno chirp-v3.5/v4/v4.5/v5 (cards upgrade to chirp-v6 on generation), Gemini 2.0-flash, and gemini-3-pro-preview (migrated to gemini-3.1-pro-preview).
  • Agent panel upgrade — the Fusion deliberation panel now uses Kimi K3 (1M context) instead of Kimi K2.6/K2.7 Code; describe catalog gained Gemini 3.8 Flash and GPT-5.6 with matching token pricing.

​
🔐 Credentials & reliability

  • Signing out now clears local credentials — the session ends and provider/relay keys, access/secret keys, tokens, MCP secret references and the local vault key are all removed from the device; base URLs and model choices stay. The button reads Sign out & clear local credentials.
  • Encrypted snapshot is always the final write — the credential snapshot is written in full first (so a reload can never lose a key), then upgraded to ciphertext, with the two writes strictly ordered so a late plaintext transaction cannot overwrite it.
  • No duplicate results — one upstream task consumed twice (in-flight request + recovery poll) no longer renders the same image twice; recovery skips cards that still have an in-flight request.

​
✨ Improvements

  • Agent runtime profiles — Settings can test and save custom describe model capabilities locally, with endpoint-scoped profiles for direct and relay runtimes.
  • Capability-aware routing — Agent calls now resolve runtime config and capability plans once, then execute with the selected profile instead of re-reading stale Settings state.
  • Claude-native agent runtime — Custom describe runtimes can use Anthropic’s native Messages API while OpenAI-compatible providers continue to cover DeepSeek/Groq/OpenRouter-style relays.
  • Provider-first model onboarding — Custom describe setup now offers provider presets plus live model discovery and manual model ID fallback, reducing the work needed to connect new model providers or relays without developer changes.

​
🛠 Fixes

  • Vision payload injection now follows the resolved runtime/profile capability, preventing non-vision custom endpoints from receiving image inputs.
  • Required probe failures no longer persist successful-looking model profiles.

​
v0.9.0 — March 2026

​
🚀 New Features

  • Creative Workshop — Staged creative project flow: Script → Storyboard → Video generation
  • Desktop Engine (local-core) — Rust-native daemon with SQLite WAL, FFmpeg, and MCP Gateway
  • Harness Engineering v2 — Tool output self-declaration with Zod schema validation
  • Confidence Calibrator — Agent self-evaluation and reliability scoring

​
✨ Improvements

  • 5-tier Unified Intent Router reduces avoidable full LLM turns for simple commands
  • DAG concurrency pipeline for parallel tool execution
  • Intent Drift Detector helps detect and recalibrate drift during long tool loops
  • Masonry Canvas virtualization for ultra-low RAM consumption

​
🛡 Security

  • Guarded mutation metadata for destructive and external-effect tools
  • Sensitive settings fields are encrypted before browser IndexedDB persistence when available
  • SQLite WAL mode for crash-safe local persistence

​
v0.8.0 — February 2026

​
🚀 New Features

  • MCP Gateway — Connect external tools via Model Context Protocol
  • ReferZone — Drag-and-drop intake for local-first asset preview and routing
  • Multi-model support — Veo 3.1, Flux Pro, Suno Chirp 5

​
✨ Improvements

  • Card-based UI replaces linear chat flow
  • IndexedDB persistence for offline card recovery
  • Batch action hooks for multi-card operations

Feature Request?

We’d love to hear from you. Submit feature requests and bug reports on GitHub.