MCA v0.2.0-alpha

MCA v0.2.0-alpha

Feature and compatibility update for the public alpha line.

Highlights

• Added configurable web search with direct public URL reading, keyword search,
  source cards, search traces, local diagnostics, and per-turn controls from the
  chat composer.
• Added research-mode controls that follow the same lightweight capsule pattern
  as reasoning mode.
• Added search provider setup for SearxNG, Brave Search, Tavily, Jina Search,
  and custom JSON search gateways, with in-app guidance for obtaining endpoints
  or API keys.
• Added the in-app API Usage Guide for OpenAI-compatible local API setup,
  trusted LAN access, supported paths, JSON replies, and SSE streaming.
• Improved local API compatibility for third-party OpenAI-compatible clients by
  preserving client-provided system messages and character-card prompts.
• Added authenticated local API vision self-test and /v1/models vision
  diagnostics without exposing generated local API keys in logs or shell
  commands.
• Improved cloud multimodal request forwarding for compatible OpenAI-style and
  Anthropic-style vision engines by preserving inline image data URLs.
• Added experimental local vision support for multimodal GGUF models with
  matching mmproj / projector files.
• Added a ModelScope recommendation bundle for MiniCPM-V 4.6 Q4 local vision:
  MCA downloads the main GGUF and matching projector, then registers and binds
  them together.
• Added local model-card actions to bind or replace mmproj / projector files,
  including projector files downloaded from remote model file lists.
• Reworked cloud inference-engine and image-generation-engine setup from small
  dialogs into full mobile pages with clearer forms and visible test results.
• Kept cloud inference engines and image-generation engines separated, including
  support for OpenAI-compatible chat, Anthropic Messages, OpenAI Images,
  DashScope Image, and custom image paths.
• Improved model-management cards so action buttons no longer cover model
  names, while local image engines and local vision bundles remain clearly
  labeled.
• Added assistant and role-card improvements, including MCA role-card
  import/export and compatibility import for common nested character-card
  data fields.
• Preserved the existing MCA chat surface and visual language while aligning
  secondary-page transitions and cleaning up menu text.

Local API Notes

Use http://127.0.0.1:11435/v1 for same-device clients.

Use http://<phone-lan-ip>:11435/v1 for another trusted device on the same
network, after enabling open port inside MCA.

Supported paths:

• GET /health
• GET /v1/models
• POST /v1/chat/completions
• GET / for the built-in web chat page with text and image attachment input

/v1/models includes MCA extension fields such as vision_ready and
vision_projector so clients can detect whether the loaded local model is ready
for image input.

The local API key is generated inside MCA and should not be shared in issues,
screenshots, videos, or release notes.

Validation

• :core:download:testDebugUnitTest
• :core:modelstore:testDebugUnitTest
• :core:engine:testDebugUnitTest
• :api:local:testDebugUnitTest
• :app:testDebugUnitTest
• :feature:modelhub:compileDebugKotlin
• :feature:chat:compileDebugKotlin
• :app:compileDebugKotlin
• :app:assembleRelease
• Release APK install and launch smoke test on a real Android device.
• Cloud vision mock smoke test with OpenAI-compatible image input enabled.
• Authenticated local API vision self-test on a real Android device with
  MiniCPM-V 4.6 Q4 and mmproj-model-f16.gguf.
• Model management smoke test covering local models, cloud engines,
  recommendations, and local vision projector setup surfaces.

Known Limits

• MCA does not include model weights, provider accounts, cloud endpoints, or API
  keys.
• Local API chat requires a local model to be loaded first.
• Same-LAN local API access should only be enabled on trusted networks.
• Cloud provider compatibility depends on each provider's exact protocol,
  route, model name, and image-input support.
• Local vision is not automatic for ordinary GGUF files. It requires a
  compatible multimodal GGUF model and matching projector, and the native runner
  must report visionReady=true.
• Local image generation remains experimental. Performance depends heavily on
  device, model bundle, resolution, step count, thermal state, and available
  memory.
• Keyword web search requires a configured search provider. The public JSON
  self-check source is only a protocol check, not a production search engine.