JiuMe

Your local-first desktop twin for personal agents.

JiuMe turns a personal agent into a small desktop companion: an always-on avatar, a one-line task box, local twin memory, mounted personal skills, and a review step before anything durable is learned.

It is intentionally local-first. You can open the app without cloud credentials, keep twin data on your machine, and decide when optional model or image providers are connected.

Why People Try It

• One-line daily loop: click the avatar, type one task, and let the agent work.
• Desktop-native feel: the everyday surface is an avatar, not another busy dashboard.
• Local twin data: profile, avatar assets, memories, permissions, and audit logs live under ~/.jiuwenswarm/jiume/.
• Review before learning: JiuMe can distill useful context into memory or personal skills, but durable writes stay user-approved.
• Built on a real agent runtime: JiuMe uses JiuwenSwarm instead of faking backend actions.

Screenshots

These screenshots are generated from the current JiuMe desktop UI renderer with a temporary local twin. No cloud provider is required.

Start From Zero

You need Python 3.11+ and uv.

git clone https://github.com/XiaoLuoLYG/jiume.git
cd jiume

uv venv --python 3.11 --seed .venv
uv pip install -e ".[test]"

.venv/bin/jiume-launch --open-setup

What should happen:

1. The setup page opens.
2. The local JiuwenSwarm runtime starts.
3. The JiuMe desktop avatar appears.
4. You can click the avatar and send a one-line task.

No API key is required for the first local run. Avatar art comes from a Codex pet package. Generate one with the local hatch-pet skill or download an existing Codex pet project, then import the folder or ZIP in Setup:

.venv/bin/jiume-setup --pet ~/.codex/pets/my-pet --enable

For a UI-only smoke test:

.venv/bin/jiume --gateway-url ""

Core Architecture

flowchart LR
  User["You"] --> Avatar["Desktop avatar"]
  Avatar --> Task["One-line task"]
  Task --> Runtime["JiuwenSwarm runtime"]
  Runtime --> Tools["Agent tools and skills"]
  Avatar --> Twin["Local twin data"]
  Twin --> Review["Review-gated memory and skills"]
  Review --> Twin

Loading

The only idea you need first: JiuMe is the friendly desktop layer; JiuwenSwarm is the agent runtime; your twin data stays local unless you explicitly connect outside services.

Project Map

jiume/                    JiuMe product layer
  desktop/                avatar window, task capsule, local desktop state
  setup/                  first-run setup and settings server
  runtime/                bridge into JiuwenSwarm
  twins/                  local profile, memory, permissions, audit data
  skills/                 personal skill catalog helpers
  personal_distillation/  review-gated learning pipeline
  avatar/                 avatar assets and manifests
jiuwenswarm/              underlying agent runtime
tests/unit_tests/jiume/   focused JiuMe tests
docs/en/JiuMe.md          practical English guide
docs/zh/JiuMe.md          practical Chinese guide

Developer Check

.venv/bin/python -m compileall -q jiume jiuwenswarm/start_services.py
.venv/bin/python -m pytest -q --no-cov tests/unit_tests/jiume
git diff --check

Learn More

• Chinese README
• JiuMe practical guide
• Contribution guide
• Security policy
• Support

Contributing

Contributions are welcome when they make the local experience clearer, safer, or more useful. Please keep user-facing claims grounded in real behavior and read CONTRIBUTING.md before opening a pull request.

License

JiuMe is released under the Apache License 2.0.