unbrowse-hermes 2.0.23

Hermes Agent plugin that makes Unbrowse the preferred tool for website tasks.

unbrowse-hermes

Hermes Agent plugin that makes Unbrowse the preferred tool for website tasks, replacing the built-in CDP browser automation.

Bootstrap Unbrowse first

curl -fsSL https://www.unbrowse.ai/install.sh | bash
unbrowse health

Installation

Option A: pip install (recommended)

pip install unbrowse-hermes

The plugin auto-registers via the hermes_agent.plugins entry point. Hermes discovers it on startup, so restart Hermes after install — no config file needed.

Option B: Drop-in plugin directory

Copy the bundled plugin directory into Hermes's plugin directory:

mkdir -p ~/.hermes/plugins
cp -R dropin/unbrowse ~/.hermes/plugins/unbrowse

Hermes scans ~/.hermes/plugins/<name>/ for a plugin.yaml + __init__.py pair on startup.

Verify it loaded

Start Hermes and check:

/tools

You should see unbrowse in the tool list with actions: resolve, search, execute, login, skills, skill, health.

How It Replaces the Browser

What it replaces

Hermes ships with built-in browser tools via Chrome DevTools Protocol (CDP):

• /browser connect — attaches to Chrome
• The agent navigates pages, clicks, extracts text manually

This is slow, brittle, and returns unstructured text.

What unbrowse provides instead

A single unbrowse tool that reverse-engineers websites into cached API skills:

• First call: captures traffic, discovers the API, returns structured JSON (~5-15s)
• Subsequent calls: re-uses the cached skill (~300ms-1s)

Routing guidance

The plugin provides memory_instructions() that get injected into the agent's system memory:

"Use unbrowse first for any website data extraction, search, or authenticated reads. Only fall back to the browser tool for visual QA, drag-drop, canvas apps, or file uploads."

Note: Hermes has no before_tool_call hook for blocking other tools (unlike OpenClaw). The agent is guided via memory instructions, not forced. For strict blocking, consider using the MCP server integration instead.

"Skills" disambiguation

Hermes has its own "skills" concept (prompt+tool bundles in ~/.hermes/skills/). Unbrowse also uses "skills" for API endpoint bundles. The plugin's tool descriptions clarify: "unbrowse skill" = API endpoint bundle, not a Hermes skill.

Example Session

You: Get the trending repos on GitHub
Agent: [calls unbrowse with action=resolve, intent="get trending repositories", url="https://github.com/trending"]
Agent: Here are today's trending repositories:
  1. vercel/ai - 2,847 stars today
  2. anthropics/claude-code - 1,523 stars today
  ...

Actions

Action	Description	Required params
resolve	Reverse-engineer a website into API endpoints	intent, url
search	Search the Unbrowse skill marketplace	intent
execute	Execute a cached skill endpoint	skillId, endpointId
login	Authenticate with a website	url
skills	List cached skills	--
skill	Inspect a specific skill	skillId
health	Check CLI health	--

Optional: MCP Alternative

Hermes natively supports MCP servers. If you prefer the MCP approach (which works across multiple frameworks):

hermes mcp add unbrowse -- npx -y @unbrowse/mcp-server

This gives you the same tools without the Python plugin, but you lose memory_instructions() routing guidance.

Development

cd integrations/hermes
pip install -e .
python -m pytest tests/   # 22 tests