Open Source Radar — November 8: An Android Agent That Uses Your Apps

Share
Open Source Radar — November 8: An Android Agent That Uses Your Apps

Today's board is agents on the move again — one that tapped into your phone's own apps, one that keeps coding after you close the terminal, one that reads a video frame by frame, and one built to make the first kind expensive to point at your binary.

Mika (Kotlin, about 7,700 stars, Apache-2.0) puts the agent on the phone itself: wangzan101/Mika opens apps, taps, types and scrolls on a hidden background screen while you keep using the phone, finding buttons by looking at the screen rather than through per-app integrations. There is no Mika server — you download the APK, pair wireless debugging once, and hand it your own model key, so tasks and history stay on the device, with a Linux sandbox per conversation for Python and shell on files you send. The demo list is the pitch: greet recruiters on a job site, find the cheapest nonstop flight, stack coupons on a takeout order and stop before paying. Created October 29, it was the number-one Kotlin repository of the day on Trendshift this week, and it is the first on-device phone agent here that is not a research demo with a waiting list.


BLA (TypeScript, about 3,200 stars, MIT) is the defence against the other side of that trend. Point qybaihe/bla at a build you ship — an Electron app, a JS bundle, a Python package, a WASM module, a safetensors file — and it plants format-valid decoys, honey metadata and deep nesting that push an AI reverse-engineering agent like REA past its own size and nesting budgets, leaving it with a truncated, partly wrong picture while your app behaves exactly as before. The published runs are specific: a 128-byte safetensors stub became 10 KB and tripped a 32-level JSON nesting limit; an Electron ASAR header went from 607 bytes to 16.9 MB against a 16 MB cap. The project is also candid about the edges — the Ghidra, Frida and JADX paths are shielded in code but not measured — so read "bankrupt" as covering the artifact formats, not the entire reverse-engineering stack.


Agent Autoloop (TypeScript, about 4,000 stars) fixes the part of coding agents nobody has: a Claude Code or Codex session dies when you close the terminal. KKKKhazix/agent-autoloop runs a small scheduling kernel on your own machine — a cron line, a new GitHub issue or a production alert becomes an event, a rule names which agent takes it and in which folder — and the agent works one round, prints its result and exits. When it reaches a decision that is yours — deleting data, restarting a service, spending money, messaging a person — it pushes a card to Slack, Discord, Feishu, DingTalk or WeCom and stops, so nothing is running and no tokens are burning while you take your time; one tap resumes the same conversation. It reports tokens and cost for today and the past seven days, and it picked up 4,000 stars in five days, which says how badly people want their agents to keep working after they walk away.


rea-video (Python, about 3,900 stars) answers a question with real search volume behind it: was this video filmed, generated or coded? JuneYaooo/rea-video cuts a clip into shots, splits the audio into voice, music and effects, then ranks hard evidence — frame cadence, sensor noise, whether the whole frame pans as one plane, AIGC tags in the metadata — above anything that merely looks off, and escalates disagreements to a vision model or the agent's own review. What comes back is a shot-by-shot verdict plus a remake plan: prompts for the generated shots, gear and lighting for the filmed ones, a Remotion project for the motion graphics. Its own eval claims 98.3% accuracy on the tuning set and 100% on recompressed clips once the agent reviews the uncertain shots — the same measure-and-verify idea behind REA, pointed at video instead of binaries.

Worth watching this week: whether Mika's on-device approach survives the Play Store's accessibility rules, and whether BLA's decoys hold once the reverse-engineering agents start budgeting for them.

Would you let an agent run your phone unattended, and what would you make it ask permission for first? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Open Source Radar — November 7: LobeHub's 84,000-Star Super-Agent

Open Source Radar — November 7: LobeHub's 84,000-Star Super-Agent

Today's board is about agents that hold down a job rather than answer a prompt: a team-chat super-agent with 84,000 stars, a Rust engine that studies any app so an agent can rebuild it, and a skill that strips the AI look out of your interface. LobeHub (TypeScript, about 84,300 stars) is the open-source platform behind LobeChat, rebuilt around what it calls super-agents: you add one to the team chat you already use, plug in any model or IM app, and it runs several tasks at once, picks up work…

Open Source Radar — November 6: VibeVoice's Hour-Long Transcript

Open Source Radar — November 6: VibeVoice's Hour-Long Transcript

Today's trending board is heavy on tooling rather than new models: Microsoft's speech family is back, two agent-facing CLIs are shipping daily fixes, and Google has quietly built a 187-skill catalog for its own cloud. VibeVoice (Python, about 56,300 stars, MIT) is Microsoft's open speech family, and the one most people will actually reach for is VibeVoice-ASR, which accepts a full 60 minutes of audio in one pass inside a 64K-token window and returns a structured transcript — who spoke, when…

Open Source Radar — November 5: A 2.78T Model on a Laptop CPU

Open Source Radar — November 5: A 2.78T Model on a Laptop CPU

Today's trending list is about plumbing rather than demos: the tools that make coding agents act like a team, keep a 2.78-trillion-parameter model off the GPU, and hand an agent data it can actually trust. Four worth a look, none of them repeat coverage from this week. gstack (TypeScript, about 139,000 stars) is Garry Tan's own Claude Code setup, published as MIT-licensed Markdown — 23 tools that reassign the agent into roles with names you'd recognise from a real org chart: a CEO that…

Open Source Radar — November 3: GLM-5.2 on a laptop

Open Source Radar — November 3: GLM-5.2 on a laptop

Today's trending board has one theme: shrinking the machine AI needs. A C engine that streams a 744B model off your SSD, a tool that strips refusals out of a model in under an hour, desktop control that skips screenshots entirely, and a git wrapper built for running ten coding agents at once. colibri (C, 46,865 stars) runs mixture-of-experts models far bigger than your RAM by keeping the always-on parts resident and reading the rest of the experts from disk as each token needs them.

Open Source Radar — November 2: 290B models on a desktop GPU

Open Source Radar — November 2: 290B models on a desktop GPU

Today's board is about putting big models on hardware you already own — one serving engine for gaming PCs, one foundation model for stock charts, and one security operations center that runs without a single API key. FreeToken (Python, 15,691 stars, Apache-2.0) is an inference engine built for the opposite of a datacenter: it runs frontier-scale mixture-of-experts models — 290B parameters and up — on a gaming PC by treating your GPU, CPU and system memory as one elastic pool, according to its…

Open Source Radar — October 30: OpenMausBot, an Open Grok Bot

Open Source Radar — October 30: OpenMausBot, an Open Grok Bot

Today's board keeps circling one question: what if your coding agent had coworkers? Two projects answer with a desktop app that hands every bot its own computer; a third gives a single session the org chart of a game studio. OpenMausBot (TypeScript, 7,848 stars, Apache-2.0) is the open-source answer to xAI's Grok Bot — a chat app where each contact in the sidebar is a real agent, Claude or Codex running under the hood, with its own personality, its own model, its own sandboxed computer and…