AIAny
Icon for item

nanoMuse: An Open-Source Personal Agent for Every Device You Own

Runs an open-source personal software agent across a user's devices to perform actions, keep readable provenance-backed memories, and coordinate via a self-hostable relay. Ships minimal end-to-end clients (mobile, desktop, web) under GPL-3.0; model-agnostic and designed for inspectability and device

Introduction

The field now treats "agents" as long-lived programs that act on accounts, devices and files; vendor products lock that capability inside cloud VMs. The core insight of this work is simple but consequential: a personal agent that holds a person's life must be inspectable, run where the person chooses, and have an open relay anyone can run. nanoMuse is a minimal, complete reference that shows how to build that open counterpart.

Key Findings
  • Minimal end-to-end reference: implements Android, iOS (TestFlight), desktop (Windows/macOS/Linux) and web clients plus a relay, all published under GPL-3.0. This demonstrates a full-stack, self-hostable architecture rather than a partial SDK.

  • Readable, provenance-backed memory: uses plain files on-device (Markdown on phones; append/change-line store on desktop) so users can read, edit and audit what the agent remembers. This design prioritizes transparency over opaque cloud-only memory stores.

  • Relay-first sync with privacy controls: the relay stores account metadata, conversation text the user opts to sync, model call ledgers (model, token counts, cost) but not user files or tool outputs; two one-tap switches control sync and opt-in model-training data sharing. This balances cross-device continuity with user control.

  • Practical resource and cost profile: example binaries are small (Android ~38 MB; desktop downloads ~256–498 MB; idle relay ≈85 MB RAM); a self-hosted relay fits on a cheap 1 vCPU/1 GB tier (estimated US$4–6/month) plus model usage costs, showing the concept is deployable outside large clouds.

Who it's for and trade-offs

Great fit if you want an auditable, self-hostable personal agent reference: researchers, privacy-conscious developers, and hobbyists who need cross-device automation with readable memory, hands-on device control, and the ability to swap model providers. It is intentionally minimal—each piece is the simplest thing that works—so it is useful as a reference or starting point rather than a polished, fully hardened commercial product.

Look elsewhere if you need a turnkey, enterprise-grade managed agent with legal SLAs, vendor-hosted compliance guarantees, or broad production skills out of the box; nanoMuse leaves safety, large-scale multi-user hosting, and advanced skill marketplaces as future work. Expect some manual setup (keys or self-hosting), limited built-in models by default, and deliberate UX trade-offs to preserve inspectability and user control.

More Items

Iteratively refines agent-generated games using a closed loop of Designer, Builder, Coding‑Native Player, and Experience‑Oriented Reviewer — collects programmatic gameplay trajectories and trajectory+visual evaluations to evolve prototypes into player-focused, product-level games; shows measurable gains on GameCraft-Bench and GameASG-Bench.

Analyzes where LLM-based agents break the evidence-to-action chain and introduces SafeActBench, a 656-case, provenance-bound benchmark and deterministic evaluator to diagnose failures in investigation, timing, single-action execution, and multi-step workflows.

Verifies and preserves trajectory-derived skill edits for LLM agents by pairing each proposed edit with replayable execution evidence and re-executing the relevant trajectory segments. Introduces Replayable Evidence Cards, a replay-based verification gate, and a Provisional Edit Ledger to retain locally supported edits across epochs for continual skill evolution.