win4r/MuseAI-Skills

MuseAI‑Skills – What the repo actually is

Short description – A snapshot of the Muse AI personal‑assistant platform (the commercial service at https://muse.ai). The repository contains:

• Markdown “skill” specifications that describe when a skill is triggered, what tools it may call, what permissions it needs and how its output is validated.
• Configuration files (home.yaml, skills.yaml), runtime‑cell launch scripts and a handful of ELF binaries (stored with Git‑LFS, not included in the source view).
• Documentation that explains the overall product architecture, connector manifests, and a static analysis report.

It is not a full open‑source implementation of Muse AI, nor a ready‑to‑run installer. The core engine is a closed‑source Linux x86‑64 binary that is only referenced, not built from the sources in this repo.

---

Who might find this useful?

• Researchers of LLM‑driven agents who want concrete examples of how a commercial AI‑assistant structures its “skills” (i.e., modular tool‑calling workflows).
• Developers building their own agent platforms and looking for a reference for:
  • Skill manifest format (SKILL.md + manifest.yaml + optional eval YAML).
  • How to separate task orchestration (wide‑research, goals, forget) from artifact handling (document, pdf, spreadsheet).
  • Permission scoping for third‑party connectors (Gmail, Google Calendar, Plaid, etc.).
• Security auditors interested in the permission model and the static analysis report included in the repo.

---

Key parts of the repository

Path	What it contains	Why it matters
opt/hatch/skills/	68 independent skill definitions (SKILL.md) plus manifest files and evaluation scenarios.	Shows the exact trigger conditions, input/output contracts, and failure handling for each capability (e.g., travel‑planning, gmail, plaid).
home/hatch/config/	home.yaml (overall agent effort config) and skills.yaml (list of enabled skills).	Gives a high‑level view of how the platform enables/disables skills and allocates compute resources.
opt/hatch/runtime-cell/	Shell scripts that start/stop the “runtime‑cell” daemon, environment setup files, and a few config snippets (skill‑scopes.conf).	Demonstrates the runtime orchestration layer that executes the skills on a Linux host.
PROJECT_ANALYSIS.md	A manually written static analysis: architecture diagram, missing pieces, risk assessment, evidence index.	Helpful for anyone who wants a quick audit of what is present vs. what is absent.
SHA256SUMS	Checksums for the non‑binary files shipped in the repo.	Allows verification that the snapshot has not been tampered with.
opt/hatch-image/	A collection of ELF binaries, npm packages and other tools packaged for the image (stored via Git‑LFS).	Indicates that the actual executable components exist but are not built from source here.

---

How the “skill” system works (high‑level)

1. Trigger – A user request matches a SKILL.md file’s trigger description (e.g., “when the user asks to book a flight”).
2. Scope check – The runtime reads the associated manifest.yaml to verify the required OAuth scopes or API keys are present.
3. Tool orchestration – The skill calls one or more connector CLIs (e.g., duffel, gmail) that are packaged in the image.
4. Result validation – Each skill includes an artifact testing section that defines how the output is inspected (PDF rendering check, placeholder scan, etc.).
5. Acceptance – If validation passes, the result is stored in the user’s “artifact library”; otherwise a failure contract is returned.

---

What’s missing / not included

• The source code of the Muse AI inference engine (the large ELF binary).
• Full root filesystem and runtime dependencies needed to actually launch the daemon.
• Any OAuth credentials or user‑specific tokens – the manifests only list required scopes.
• Automated tests or CI pipelines; the eval YAML files are just data for manual testing.

---

Quick start for reading only (no binaries)

# Skip downloading the large LFS binaries
GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/win4r/MuseAI-Skills.git
cd MuseAI-Skills
# Browse the skill specifications
ls opt/hatch/skills/gmail
cat opt/hatch/skills/gmail/SKILL.md

---

Bottom line

MuseAI‑Skills is a documentation‑focused snapshot of a commercial AI‑assistant’s modular skill set. It is valuable as a reference for how a production‑grade LLM‑agent structures its tool‑calling workflows, permission manifests, and artifact validation, but it does not provide a runnable open‑source implementation of Muse AI itself.