AndroidHarness brings installable agent plugins to your phone

AndroidHarness 3.2 turns its phone-based coding agent into a host for installable plugin bundles. The release announces npm, GitHub and local-file installation, alongside /btw side questions and a Hugging Face search hub for on-device models.

For developers working from a phone, the useful change is packaging a repeatable coding setup rather than rebuilding it inside every conversation. AndroidHarness is a standalone native Android app, not a model or an OpenCode/Pi extension. Its earlier multiple-project support is separate from this update.

Bring a coding setup, not an arbitrary desktop package

The current English plugin guide describes bundles that contribute instructions, skills, MCP server definitions and optional JavaScript hooks. Installation lives under Settings > Coding > Plugins, with project or global scope and a preview of the actual contents. Changes apply to the next Code run; an active turn keeps its captured configuration.

Compatibility matters. A bundle needs AndroidHarness’s plugin.json format. Packages written for another harness are not automatically compatible, even when they come from npm. The installer does not install dependencies or execute npm lifecycle scripts. GitHub sources resolve to an immutable commit, while npm archives are checked against their published integrity hash, according to the guide.

Installing a plugin also does not start its MCP server or grant execution permission. Server startup and executable hooks retain separate approval. The documented hook runtime has no filesystem or network bridge and requires a supported Android System WebView; skill-only bundles do not need that JavaScript runtime. These are the project’s documented controls, not an independent security assessment.

Ask a side question while work continues

The 3.2 changelog says /btw reuses the main task’s latest request for a quick question while the agent continues. It may use read-only tools, but the side question and answer do not enter the main conversation. Token usage still counts toward the chat’s cost. Without a running task, it answers from the transcript instead.

The same release announces resumable local-model downloads and Hugging Face search. Current documentation describes suggestions based on the phone’s chip, memory and free storage. That is a discovery aid, not a guarantee that every listed model will run well. Local inference still needs a compatible 64-bit device and sufficient resources.

Availability and evidence

The GitHub release includes an APK. The project supports Android 8.0 and newer and also links Google Play, though this review did not confirm the store’s current version. Cloud models use configured keys or eligible connected accounts; the built-in keyless provider has its own displayed limits. Local models can run offline after setup, while network tools need connectivity.

The setup guide requests storage access, including All files access on Android 11 and newer for real filesystem paths. Cloud-backed tasks can send project context to the chosen model provider.

There is a documentation mismatch worth preserving: the 3.2 release announcement lists these additions, but the README and changelog in its Git tag still describe 3.1. The detailed behavior above comes from the separately pinned current English 3.2 documentation. AITrending reviewed both sets of sources but did not install the APK or verify feature parity on a phone.