dev-boffin-io/mind-forge-pro
Watch
1
0
Fork
You've already forked mind-forge-pro
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
No description
27 commits 1 branch 0 tags 157 KiB
  • Dart 99.1%
  • Kotlin 0.9%
Find a file
Exact
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Boffin 8f413d0e3f
Some checks failed
Build Mind-Forge Pro APK / Build release APK (arm64) (push) Has been cancelled
Details
feat: add custom OpenAI-compatible and Ollama endpoint support 
Add a persistent Settings tab so users can point Mind-Forge at any
OpenAI-compatible API or Ollama host instead of (or alongside) the
local llama.cpp engine.

- app_settings.dart: shared_preferences-backed singleton storing the
  chat backend (local/openai/ollama), OpenAI base URL + API key +
  model, and Ollama host + model. Blank fields gracefully fall back to
  https://api.openai.com/v1 and http://localhost:11434.
- api_client.dart: RemoteApiClient abstraction with OpenAI-compatible
  /chat/completions and Ollama /api/chat clients. Clients read
  configuration from AppSettings at every call, so edits apply
  immediately without restarting. OpenAI key is optional (no auth
  header for local servers).
- chat_logic.dart: send() picks the active RemoteApiClient from
  settings and routes turns through it; the agentic retry loop works
  for both remote and local backends.
- main.dart: new Settings tab (backend selector + endpoint fields),
  async settings load before runApp, and a chat banner showing the
  active backend / prompting to load a model for local mode.
- pubspec.yaml: add shared_preferences and http dependencies.
2026-09-19 12:19:31 +00:00
.github/workflows Fix: Resolve Android SDK setup failure in CI workflow 2026-09-16 19:59:02 +00:00
android Fix Request naming collision, bump compileSdk to 36 2026-09-03 06:23:49 +00:00
lib feat: add custom OpenAI-compatible and Ollama endpoint support 2026-09-19 12:19:31 +00:00
.gitignore Migrate to llama_cpp_dart 0.9.x — prebuilt AAR instead of from-source NDK build 2026-09-03 03:21:46 +00:00
LICENSE
pubspec.yaml feat: add custom OpenAI-compatible and Ollama endpoint support 2026-09-19 12:19:31 +00:00
README.md Migrate to llama_cpp_dart 0.9.x — prebuilt AAR instead of from-source NDK build 2026-09-03 03:21:46 +00:00

README.md

Mind-Forge Pro

An offline, on-device "second brain" AI agent for Android. No Termux, no external Ollama server — the app embeds llama.cpp natively via Dart FFI and serves itself over a local HTTP API.

How it works

  • Chat — talk to a locally-loaded .gguf model. Every turn is grounded with a lightweight TF-IDF retrieval pass over your own stored notes (lib/memory_agent.dart) before it reaches the model.
  • Memory DB — browse, inspect, and delete anything the app has remembered (SQLite, on-device only).
  • Server Config — pick a .gguf file from local storage and start a local API server (http://127.0.0.1:<port>/api/generate) that any other app or script on the device can call.

Architecture

Layer Tech
UI Flutter (Material 3)
Inference llama_cpp_dart 0.9.x, off the UI thread via LlamaEngine/EngineSession
Local server shelf + shelf_router
Memory / RAG sqflite + a pure-Dart TF-IDF keyword ranker
Native actions a MethodChannel the model can trigger via <ACTION: X> tags

Building

CI (.github/workflows/build-apk.yml) does everything needed for a sideloadable release APK — no production keystore required:

  1. Downloads llama_cpp_dart's prebuilt Android AAR (CPU + mtmd, arm64-v8a) from its GitHub Releases, matching the exact version pinned in pubspec.yaml, into android/app/libs/llama-cpp-dart.aar. No native compilation happens in this project at all — 0.9.x ships binaries built against its own pinned llama.cpp commit, which is also why the earlier from-source NDK build (against an unpinned llama.cpp clone) kept hitting ABI mismatches.
  2. Runs flutter build apk --release --target-platform android-arm64 with R8/ProGuard shrinking and native debug-symbol stripping enabled (android/app/build.gradle).
  3. Signs the release build with the standard Android debug keystore, cached across runs so every build shares the same signing key — that's what lets you install an updated APK over a previous one without uninstalling first. (This is fine for sideloading; it is not a Play Store–ready signature.)
  4. Uploads the finished APK as a workflow artifact.

To build locally instead, download llama-cpp-dart.aar yourself (see the package's install instructions) into android/app/libs/ before running flutter build apk.

Status

Early scaffold — chat, memory, and server-management UI are wired up; model loading and generation go through llama_cpp_dart's LlamaEngine/ EngineSession API (0.9.x line — a prerelease, still pre-1.0 with occasional breaking changes). Verify current API shape against whatever llama_cpp_dart version is pinned in pubspec.yaml before assuming any snippet here still matches.