I work from a phone more often than I would like to admit. Most mobile AI apps I tried were front-ends: one provider, a prompt box, maybe a document picker. As soon as I needed a shell on a VPS, a file on the phone, an MCP server at home, or the ability to tap something in another app, I was back at a computer.
So I built an AI app for Android. ForgeMind is a native Kotlin + Jetpack Compose app. I wanted a workstation on the device, not another thin chat client: cloud and local models, SSH (including Cloudflare Access), MCP in both directions, files and a shell, and a way to read and operate the screen of other apps.
This article is about how it is put together, what I traded away, and how to try it. The repository is a new public release (mid-November 2026). ForgeMind is 0.2.1-alpha.
A dual-contour hub on Android
The short version of the architecture is in the README. The idea is a dual contour — external connections and on-device ones — without shipping a multi-gigabyte Linux container inside the APK.
Outside the phone, the app speaks five native LLM wire protocols: OpenAI-compatible /chat/completions (SSE, with a non-SSE fallback), OpenAI /responses, Anthropic /v1/messages, Gemini streamGenerateContent, and Ollama's native /api/chat. There are twelve cloud presets (OpenAI, Anthropic, Gemini, OpenRouter, DeepSeek, xAI Grok, Groq, Mistral, Venice AI, NVIDIA NIM, Cloudflare Workers AI, Hugging Face), plus arbitrary custom HTTP gateways (six auth modes, custom headers and body, optional insecure TLS for LAN, DNS-over-HTTPS fallback), plus a LAN scanner. You can probe a single IP:port in about a second, or scan the /24 for servers such as Ollama, LM Studio, llama.cpp, KoboldCpp, vLLM, SGLang, LocalAI, Jan, GPT4All, OpenWebUI, and LiteLLM.
SSH and SFTP use JSch, with password or key auth, over a public IP, a VPN, or Tailscale. For Zero Trust I implemented Cloudflare Access SSH over wss://: the app sends CF-Access-Client-Id and CF-Access-Client-Secret and does not need a cloudflared binary on Android, and the server does not need an open SSH port.
MCP is bidirectional. As a client, ForgeMind supports Streamable HTTP (POST /mcp), HTTP+SSE, and SSH_STDIO (it can spawn npx, uvx, or python3 MCP servers on a remote host or in Termux). As a host, a built-in Streamable HTTP server listens on 127.0.0.1:8765 (/mcp and /health, Bearer token, toggles for write/shell) so a local agent, Termux, adb forward, or a tunnel can call the phone's tools.
On the device, four file environments share a dual-pane manager: the app sandbox, shared phone storage, Storage Access Framework folders, and SFTP. The Sora editor shows unified diffs and takes a rollback snapshot before an edit. A persistent /system/bin/sh keeps its working directory if SSH drops. Termux is a first-class contour (127.0.0.1:2222 and RUN_COMMAND). There is an 8-window "Turbo" agent browser, separate from the ordinary in-app browser, and a floating chat overlay.
The part I get asked about most is screen control. After you opt in to the Accessibility service, ForgeMind can read the visible text and controls of other apps as a tagged tree ([#1]…, windows [W1]…, including split screen, floating windows, and the IME), then tap, type, or scroll by tag, including a batch of actions in one call. Screenshots are on demand (Android 11+), not a background recording. Android still hides password fields, and FLAG_SECURE still blocks screenshots. I call this out because "the agent can use your phone" should not sound magical.
Under the chat loop there are 56 built-in tools (29 core + 27 in-app levers), plus whatever an attached MCP server adds. For small local GGUF models the schema is compacted automatically, and the parser understands seven inline tool-call syntaxes. The model picker can auto-detect a loaded context window (LM Studio, llama.cpp, Ollama, /v1/models) and override it with one tap.
Tradeoffs I accepted. This is an alpha. The first public tag was v0.1.0-alpha on 13 November 2026; v0.2.1-alpha followed a day later, mainly to stop the model picker crashing on aggregator catalogs with more than a thousand models. Accessibility is powerful and easy to get wrong on OEM skins (the README already mentions HyperOS / MIUI overlays). I encrypt secrets on the device with AES-256-GCM and do not ship API keys in the repo, but you are still putting a tool-using agent on a phone — the approval toggles and destructive-shell intercepts are there for a reason. The project is GPL-3.0; I retain the right to ship official builds under a separate commercial license.
Requirements: Android 10+ (API 29, compile SDK 36), Kotlin 2.2.21. I currently ship a prerelease APK from GitHub Releases (ForgeMind-v0.2.1-alpha.apk). You can also add https://github.com/DIIIUA/ForgeMind in Obtainium with Include prereleases enabled, or build from source:
git clone https://github.com/DIIIUA/ForgeMind.git
cd ForgeMind
./gradlew testDebugUnitTest assembleDebug
That Gradle invocation is also how I run the 96 unit tests. After install: Settings → AI Providers for a cloud key or a custom endpoint; Find Local AIs in Network for LAN servers; Environments for Termux or Cloudflare SSH; MCP Servers for client transports or the on-device host.
A live portal and the screenshots used in the README are at diiiua.github.io/ForgeMind.
What is still rough
ForgeMind is alpha software. I would not treat it as a daily driver without watching the first tool calls yourself, especially anything that writes files, runs a shell, or drives the Accessibility service. The APK is a GitHub prerelease, not a store build.
What I would like to hear back
How reliable tool calling is on the local models you actually run; whether LAN discovery finds your stack; how the Window-Tag actions behave on your OEM skin; and which parts feel unfinished for a post-alpha release.
Issues and pull requests are welcome. Please do not open a PR that includes API keys, hostnames, or device identifiers.
Links
- ForgeMind (Kotlin, Android, GPL-3.0, v0.2.1-alpha): github.com/DIIIUA/ForgeMind · Releases · Website
I am Anton (DIIIUA). If you try the app, I am more interested in a failed setup or a wrong tool call than in a compliment.
Top comments (0)