How to use the OpenCLI Browser Control agent skill (opencli-browser)

What is an agent skill and what does this one let the agent do?

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent loads when a task matches the skill's description. This skill's description is: drive Chrome with OpenCLI to inspect pages, interact with logged-in sites, capture API evidence, and manage native tabs, groups and windows.

Once installed, the agent can open pages, read their content and interactive elements, click, type and fill forms, take screenshots, and inspect the network requests a page makes. It can work inside sessions that reuse your logged-in Chrome profile, and it can group, rename and move Chrome tabs and windows through Chrome's native APIs.

• Inspect pages: read URL, interactive element references and page data
• Interact: click, fill, type, select, check, upload, drag, keys and dialogs
• Capture network activity and find candidate API endpoints
• Manage native Chrome tabs, tab groups and windows

How do you install the opencli-browser skill?

Download the ZIP for this skill from this page and unzip it. The skill ships as a single file, SKILL.md. Put the unzipped skill folder in one of these locations:

For a personal install across projects, place it in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. Alternatively, in the Claude apps you can upload the skill under Settings → Capabilities → Skills.

The skill only teaches commands. You also need the OpenCLI command-line tool (opencli) installed and connected to Chrome, and Google Chrome with access to the profile you want to use.

• Personal: ~/.claude/skills/opencli-browser/
• Project: .claude/skills/opencli-browser/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload the skill

How do you trigger the skill with natural requests?

You do not need to name the skill. The agent chooses it when your request matches its description. Ask for a task in browser terms: the page, the site you are logged into, the requests a page makes, or the tabs you have open.

The skill needs Bash tool access, because it runs opencli commands. If the agent cannot run commands, it cannot use the browser.

• Inspect the signup page on localhost:3000 and tell me what the form fields are called.
• Open my logged-in project board and fill in the new ticket titled Login bug.
• Find which API endpoints the pricing page calls when I click Compare plans.
• Group my open research tabs into a group called Research and make it blue.

What does a basic browser session look like?

Page commands require a session name. Reuse the same name for the whole task, because different sessions keep their page work separate. Use --profile <name> if you need to pick a specific Chrome profile. The skill teaches the agent to discover commands with opencli list <query> and to read exact arguments with help output.

Default output is one JSON value shaped like { "ok": true, "data": ... }. Failures return ok set to false, an error object with code, message and hint, and a nonzero exit code. When error.outcome is unknown, the agent is instructed to inspect the affected state before retrying a write.

• Use the same --session name across one task
• Run opencli doctor when diagnosing a connection problem
• Use opencli browser <command> --help -f json for exact arguments

How does the agent find and act on page elements?

The state command returns the URL and interactive element refs. Interaction targets accept a ref, a CSS selector, or semantic flags such as --role, --name, --label, --text and --testid. If several elements match on purpose, --nth resolves them. If a ref goes stale, the agent reads fresh state or locates the element again.

fill replaces a field and verifies its value, while type types text. Tab commands manage tabs within a session. A bound session attaches to your current user tab: it can navigate but cannot create, select or close tabs. unbind releases the binding without closing your tab.

• state --diagnostics adds page title and available initial-state globals
• state --source ax selects the accessibility tree
• find locates elements before acting
• frames lists frame targets; eval --frame evaluates inside a chosen frame

How do you capture API evidence from a page?

The network command returns captured request keys and response shapes. Use --detail with a request key to read one cached request, --filter to select shapes, and --raw to expose response bodies. The --static option adds candidates parsed from loaded scripts, with source locations and parameter names.

Static candidates are not executed or proven APIs. Check analysisComplete, script errors, omitted scripts and networkSource before drawing conclusions. An empty live capture can reuse a recent cache, and the reported provenance shows which happened.

• network: list captured request keys and response shapes
• network --detail <RequestKey>: read one cached request
• network --static: add script-derived candidates, not confirmed APIs

How do you manage Chrome tabs, groups and windows natively?

Native calls need no session and do not attach a page. Use describe to list supported methods, signatures, live permissions and official documentation. Use the numeric IDs that Chrome returns, replacing the example IDs in the commands below.

Tabs and windows created natively persist. Changes that touch active OpenCLI session tabs or managed groups return managed_resource_conflict. Ordinary groups can be created, renamed, recolored, moved and dissolved, and tabs.ungroup dissolves a group.

• describe lists all native methods; describe tabs.group shows one method's details
• Ordinary groups can be renamed, recolored and moved
• Session-managed tabs and groups are protected from native changes