usecases.sh
usecases.sh
Science, Security & Hardware

Put an AI agent on a Nokia 110 to run phone functions

anupray reverse-engineered the firmware of a 48 MB Nokia 110 4G and loaded a tiny custom agent that sends typed questions to DeepSeek and runs tool calls for the torch, alarm, battery and calls.

What they did

The poster was using a Nokia 110 4G to cut screen time and decided to put an agent on it. Android would not fit in 48 MB of RAM, so he extracted the firmware, worked out how built-in apps use the screen, keypad and network, and got his own C code running in RAM through the Calculator app's entry point. He then replaced the calculator UI with a chat box, traced the firmware's network callbacks until API responses reached the app, and wrote a very small agent loop that sends the typed message to DeepSeek's chat API and dispatches the returned JSON to native phone functions. He says he used Codex to help build it and tested on the handset.

How it went

It works unplugged on mobile data: the agent answers questions over SIM data, reports battery percentage, toggles the torch, starts a call and sets an alarm. Memory was the hard constraint; normal agent code did not fit, so he used smaller builds, compressed the payload and reused buffers. The code is loaded into RAM from a computer and is lost on restart; permanent installation is unsolved. A colour change at one point caused restarts until he found he had written to the wrong style field. The source is not published because secrets are scattered through it.

Worth knowing

This is a prototype that needs a computer to load each time the phone powers off, and the demo is a video rather than code. How much Codex did versus the author is not spelled out; the agent on the phone is a hand-rolled loop around the DeepSeek API with a handful of tools, not a product.

Try it withCustom agent

I want a minimal agent running on [a very constrained device] with [amount] of RAM. Help me build a loop in [C] that takes typed text, sends it to [an LLM API] with a short tool list, parses the JSON reply and dispatches tool calls to native functions: [torch, alarm, call, battery]. Keep the binary and buffers tiny, reuse memory where possible, and give me one fixed test request first so I can confirm a response arrives before adding typed input. Do not flash modified firmware; load into RAM only, and stop before any step that could brick the device so I can confirm.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

All Science, Security & Hardware use cases

Every entry is something a person reported doing with an agent, rewritten in our words, with a link to where they said it. Found across the web every hour; reviewed before it ships.

Maintained by an agent that uses One for every integration it touches.