agent-plugins

agent-plugins packages Agent Skills, Model Context Protocol configuration, and client extension files with a Python distribution. Each installed wheel carries the agent files that match its code version.

The package implements the portable Agent Plugins directory format and supports Python 3.10 through 3.14.

Ship an Agent Plugin

Keep one Agent Plugin tree beside the code it documents:

my-project/
├── plugin.json
├── skills/
│   └── use-my-package/
│       ├── SKILL.md
│       ├── agents/
│       ├── references/
│       └── scripts/
├── mcp.json
└── packages/
    └── python/
        └── pyproject.toml

skills/ contains Agent Skills. mcp.json contains Model Context Protocol server configuration when the plugin provides MCP servers.

Create plugin.json at the plugin root:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "my-project"
}

Configure the Python package to wrap uv_build:

[build-system]
requires = ["agent-plugins==0.1.0", "uv_build==0.12.2"]
build-backend = "agent_plugins.build.uv_build"

[tool.agent-plugins]
root = "../.."

root is relative to pyproject.toml. The build selects plugin.json, the complete skills/ tree, and mcp.json when present.

Build, install, and locate the packaged plugin:

uv build packages/python --out-dir dist
python -m pip install "agent-plugins==0.1.0" dist/my_package-*.whl
agent-plugins locate my-package

/.../site-packages/my_package-1.2.3.agent-plugin

The wheel contains the plugin directory and an agent_plugins.json marker in the distribution metadata. A source distribution stages the same files for a reproducible wheel rebuild. An editable install points the marker at the authored plugin tree.

Use Hatchling

Keep the plugin settings and select the Hatchling adapter:

[build-system]
requires = ["agent-plugins==0.1.0", "hatchling==1.31.0"]
build-backend = "agent_plugins.build.hatchling"

[tool.agent-plugins]
root = "../.."

Include other plugin files

Add root-relative patterns for executables or client extensions:

[tool.agent-plugins]
root = "../.."
include = ["bin/**", "com.example.client/**"]

Each pattern must stay within the plugin root and match at least one file.

Inspect an installed plugin

Add agent-plugins to the runtime dependencies of Python code that calls the inspection API:

[project]
dependencies = ["agent-plugins==0.1.0"]

import agent_plugins as ap

plugin = ap.locate("my-package")

print(plugin.manifest.name)
print(plugin.manifest.path)

for skill in plugin.skills:
    print(skill.path)
    print(skill / "SKILL.md")

if mcp := plugin.mcp:
    for name, server in mcp.servers.items():
        print(name, server)

ap.locate() accepts the distribution name used by pip. ap.installed() returns every discovered Agent Plugin keyed by distribution name.

Manifest, MCP, and skill documents load on first parsed-field access and cache their result. Call ap.locate() again to read a fresh snapshot. Invalid documents raise ap.ValidationError.

MCP servers are frozen ap.StdioServer, ap.StreamableHTTPServer, or ap.SSEServer values. Their fields preserve placeholders such as ${PLUGIN_ROOT} for the agent client to resolve.

Inspect a build plan

Preview the files selected by [tool.agent-plugins] before building:

agent-plugins plan packages/python

Use --json for machine-readable output. Python build integrations can consume the same plan:

import agent_plugins as ap

plan = ap.build_plan("packages/python")
for file in plan.files:
    print(file.source, "->", file.target)

List every plugin visible in the current environment:

agent-plugins list --json

Develop and release

From a repository checkout, install the locked development environment and run the local checks:

uv sync --locked
uv run ruff format --check src tests
uv run ruff check src tests
uv run ty check
uv run pyrefly check
uv run pytest -q
./scripts/build-dist.sh

Prepare the version and lockfile in a release pull request:

uv version --bump patch

After the release commit reaches main and its push CI passes, update the local branch and start the tag-driven release:

git pull --ff-only origin main
./scripts/release.sh --dry-run
./scripts/release.sh

Before the first tag, configure PyPI to trust .github/workflows/publish.yml through the repository's pypi environment. PyPI documents the setup in Adding a Trusted Publisher.