---

Installation

If you're using coding agent like Claude Code or Codex, just send it the following prompt to install automatically:

Fetch the installation guide and follow it: https://raw.githubusercontent.com/SwanHubX/SwanLab-Skill/main/README.md

For manual installation, the recommended way is a global install:

npx skills add SwanHubX/SwanLab-Skill -g -y

You can also view the skill page on ModelScope or SkillHub.

Note: -g installs at user level and -y skips interactive confirmation. Nothing is written into the current project — the skill is downloaded to .agents/skills under your home directory, with symlinks created for Claude Code, Codex, and other CLIs that support Agent Skills. Multiple agent CLIs under the same user can share this single copy.

Then log in to SwanLab:

pip install swanlab
swanlab login          # paste your API key from https://swanlab.cn
swanlab ping           # (optional) check connectivity
swanlab verify         # (optional) validate credentials

To update the skill later:

npx skills update swanlab-skill -g -y

Contents

skills/
└── swanlab-skill/
    ├── SKILL.md
    ├── references/
    │   ├── SDK_QUICKSTART.md
    │   ├── CLI_REFERENCE.md
    │   └── SWANLAB_CONCEPTS.md
    └── scripts/
        ├── plot_metrics.py
        └── runs_benchmark.py

File	Purpose
SKILL.md	Lightweight entry point and task router — tells the agent whether to read the SDK or CLI reference
references/SDK_QUICKSTART.md	Tracking code quickstart: swanlab.init / log / finish and media logging (image / audio / text)
references/CLI_REFERENCE.md	Query metrics, logs, summaries, columns, and media, and filter experiments via swanlab api
references/SWANLAB_CONCEPTS.md	Data model, terminology, and path convention (user/project/run_id)
scripts/plot_metrics.py	Line chart of a single experiment's scalar metrics
scripts/runs_benchmark.py	Compare the same metric across experiments (normalization + best-run ranking)

Capabilities

Capability	How	Entry point
Log metrics & rich media (image / audio / text) during training and fine-tuning	Python SDK	references/SDK_QUICKSTART.md
Inspect a run's metrics, logs, summaries, columns, and media	swanlab api run ... CLI	references/CLI_REFERENCE.md
List / filter experiments by config or summary conditions	swanlab api run filter CLI	references/CLI_REFERENCE.md
Manage projects, self-hosted users, and other resources	swanlab api project / user CLI	references/CLI_REFERENCE.md
Understand the data model & terminology before querying	—	references/SWANLAB_CONCEPTS.md
Plot a single experiment's scalar metrics as a line chart	Helper script	scripts/plot_metrics.py user/proj/run -k loss,acc -o chart.png
Compare the same metric across multiple experiments (normalization + best-run ranking)	Helper script	scripts/runs_benchmark.py user/proj/r1 user/proj/r2 -k loss --direction lower

Both helper scripts accept --data file.json to render from previously saved query output, and require swanlab login (or --api-key / --host).

Run Modes

swanlab.init(mode=...) controls where data goes:

Mode	Local Storage	Cloud Upload	Use Case
online	Yes	Yes	Normal cloud usage. Requires login.
local	Yes	No	Air-gapped / no account needed.
offline	Yes	No	Save locally, upload to cloud later via swanlab sync.
disabled	No	No	Completely disable all logging.

Packaging

Build the release archive:

make package

The generated dist/swanlab-skill.zip extracts to:

SKILL.md
references/
scripts/

So it can be extracted directly into a skill directory such as .claude/skills/swanlab-skill/.

License

MIT License.

Copyright (c) 2026 Emotion Machine (Beijing) Technology Co., Ltd.