Load agent skills and code-mode guidance

The pyagentbrowser wheel registers an importable marimo code-mode capability and carries an Agent Plugin with Python-specific browser instructions. It also embeds the skills from its pinned agent-browser engine for hosts that consume the native engine's instruction set.

Discover pyagentbrowser in marimo code mode

Marimo 0.24.0 or later discovers installed code-mode capabilities through Python package entry points. Install pyagentbrowser in the notebook environment, then inspect the registered module:

import marimo._code_mode as cm

print(cm.capabilities()["pyagentbrowser"])

import agentbrowser.agent as browser_agent

help(browser_agent)

help(browser_agent) renders a short SDK workflow and the paths to the Agent Plugin resources installed with the same pyagentbrowser version.

Use a named module-owned connection when browser work spans code-mode calls:

browser = browser_agent.connect("research")
browser.open("https://example.com")
page = browser.observe()
print(page.origin)
print(page.text)

# In a later code-mode call
browser = browser_agent.connect("research")

# After the final call
browser_agent.disconnect("research")

Use browser_agent.connections() to inspect every retained name and controller. Call browser_agent.disconnect_all() after a task that created several named connections.

Pass session=agentbrowser.SessionOptions(...) on the first connect() call when the task needs domain restrictions, confirmation policy, or another session option. Supplying different options for the same live connection raises ValueError with the disconnect() recovery call. If user code closes the controller directly, the next optionless connect() preserves those session options. disconnect() is the explicit boundary for forgetting them.

plugin = browser_agent.agent_plugin()
skill = browser_agent.agent_skill()

print(plugin)
print(skill.file("SKILL.md"))
instructions = skill.body

The packaged pyagentbrowser skill uses Browser, AsyncBrowser, typed namespaces, and browser.native directly. These Python operations share the embedded native action service with the upstream engine.

Read the embedded native instruction files

An agent skill is a Markdown instruction set with optional supporting files. The embedded native skills let an application supply version-matched upstream guidance to an agent.

import agentbrowser.skills as skills

print(skills.available())
print(skills.read("core"))

The core skill teaches the engine's browser interaction workflow. read(name) returns its SKILL.md content.

Inspect supporting parts

Function	Result
available(*, include_hidden=False)	Skill names
list(*, full=False, include_hidden=False)	Skill metadata, optionally with loaded files
get(name, *, full=False)	One Skill
parts(name)	SkillPart metadata for every file
part(name, path)	One SkillFile
read(name, path="SKILL.md")	One file's text
markdown(name, *, full=False)	Main instructions or combined skill Markdown

Use the main file when the agent needs the normal browser workflow. Load individual parts when a task reaches the capability they describe. Use full=True when the host needs to package the complete instruction tree in one operation.

Skill paths reject absolute paths, traversal, empty components, and malformed relative paths. Missing skills or files raise KeyError.

Skill records its name, frontmatter description, main content, part metadata, loaded files, and hidden state. SkillPart records a relative path and kind. SkillFile records a relative path and text content.