LlmBench is an early prototype evolving quickly. This page follows the released state on the repository's default branch rather than older design plans.
Completed#
Hardened provider WebViews, added a small LRU pool, pauses inactive views, and evicts them under memory pressure while provider-owned session state persists.
Moved portable provider, model, profile, and prompt logic into the Kotlin Multiplatform shared module.
Added provider file uploads through Android's file picker plus share-sheet routing for text, images, and application files.
Added direct Gemini, OpenAI, and Claude transports plus OpenRouter Free and AIHubMix gateway adapters.
Added provider-scoped tweaks/userscripts, privacy-safe diagnostics, and generating/unread tracking for the first supported web providers.
Documented full and partial WebView support, including cases where embedded sign-in or uploads still require verification.
Normalized the app identity, removed generated/build-machine files, and added build/lint checks.
Planned#
Verify newer WebView providers
Finish embedded sign-in, upload-flow, and activity-probe verification for Qwen, Microsoft Copilot, Z.ai, Grok, Character.AI, Venice, and Meta AI.
Prompt and skill workflows
Add a local Markdown prompt vault plus editable/importable SKILL.md assets with safe capability gating.
First-party document and code tools
Bring selected Docbench and Codebench validation, repair, tokenizer, and hidden-text inspection features into compatible prompt/file workflows.
Modernize chat and navigation UX
Continue Material 3 Adaptive and Compose performance work, including large-screen list-detail layouts.
Widgets and notification replies
Add provider-aware chat/message widgets and notification replies where the transport can safely send in the background.
Security and privacy release gates
Keep sensitive storage, backup rules, WebView isolation, tool permissions, redacted previews, safe exports, and regression tests as cross-cutting release requirements.
How to contribute#
Useful contributions include provider verification, shared KMP improvements, adapter maintenance, prompt/skill tooling, security testing, and measured UI performance work.
Check the repository's issue tracker and pull requests for the freshest implementation work before starting a larger change.