EV Terminal — A Local AI Agent on Android

The project

EV Terminal is a small Android agent built around a deliberately narrow split of responsibility:

• the model handles language understanding and synthesis;
• the harness decides when a tool is needed;
• deterministic tools perform operations that should be correct;
• the runtime records what happened and releases model resources after the task.

The app runs Qwen3.5-0.8B locally through llama.cpp. It does not require an Ollama server, a permanent cloud connection, or a model that stays loaded in memory all the time.

Why it is interesting

The interesting experiment is not “how smart is a 0.8B model?” It is whether a small model becomes useful when the surrounding system gives it a clear job.

The request path is:

snippet.txttext
EditCopy
1ChatFragment
2  -> AgentRunner
3    -> ModelSupervisor
4      -> Qwen3.5-0.8B / llama.cpp
5    -> ToolRegistry
6      -> math, time, weather, web, mail, location

For a tool task, EV routes the request, executes the allowed tool, feeds the result back to the model, and records the lifecycle in Console. This keeps model output away from operations that benefit from deterministic code.

The lifecycle

The model is downloaded separately from the APK. On first setup, the app resumes and verifies a roughly 563 MB model package before storing it in private app storage.

During a task, the runtime can:

1. load the local model;
2. understand the request and choose an allowed route;
3. run a deterministic tool when one is required;
4. synthesize the result for Chat;
5. unload the model and restore memory for the next task.

This is a useful pattern for local agents: model intelligence is treated as a task-scoped resource, not a permanently running service.

Tools in the repository

• Math: expressions, derivatives, antiderivatives, definite integrals, and limits
• Time: date, time, and timezone information
• Weather: network-backed weather lookup
• Web: network-backed search
• Mail: mailbox access
• Location: device location access

The calculus path also renders successful results as LaTeX locally through KaTeX.

A real test request

The repository gives a concrete end-to-end example:

snippet.txttext
EditCopy
1Chat:    differentiate x^2 sin(x)
2Console: MODEL LOAD -> MATH SUCCESS -> MODEL UNLOAD
3Chat:    2x sin(x) + x^2 cos(x)
4Console: RAM RESTORED -> TASK COMPLETE

The important observation is the boundary. The model can translate a natural request into a tool operation, but the calculus engine owns the derivative and the runtime makes the lifecycle visible.

Reproduce the project checks

From the repository root:

snippet.shbash
EditCopy
1JAVA_HOME=/usr/lib/jvm/java-17-temurin-jdk ./gradlew testDebugUnitTest assembleDebug --no-daemon

The debug APK is generated under app/build/outputs/apk/debug/.

Limits of this note

The first model download still needs network access. Weather and web depend on network availability, while mail and location depend on device access. The repository also calls out the next missing proof point: a real 15–20 second device recording showing one complete local-agent task.

This is a project note, not a performance ranking. The useful question is whether the architecture makes local AI more practical for the work you want to do.