# Automate Browser Workflows with Agent Browser

Manual browser tasks slow research, testing, and operations work. This skill gives Claude, Codex, and Claude Code repeatable browser sessions with snapshots, actions, screenshots, and videos.

## Install

```bash
npx skillstore add inference-skills/agent-browser
```

## Metadata

- Status: approved
- Slug: inference-skills-agent-browser
- Skillstore revision: r1
- Version status: missing
- Tree hash: 54409a214f23c5e7665eabf253f540985e236a5ffa0ee5c9f2da39ada320d880
- Author: inference-skills
- GitHub username: inference-skills
- License: MIT
- Repository: https://github.com/inference-skills/skills/tree/main/tools/utilities/agent-browser/
- Ref: a06681402992ceae98ba04d54cfd4ab004862696
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network, filesystem
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/inference-skills-agent-browser
- Manifest: https://skillstore.pages.dev/api/skills/inference-skills-agent-browser/manifest

## Capabilities

- Open URLs and return compact page snapshots with @e element references.
- Click, fill, select, check, drag, upload, scroll, wait, go back, and navigate by URL.
- Capture viewport screenshots, full-page screenshots, and optional session videos.
- Execute JavaScript on the current page and return the result.
- Configure viewport size, user agent, HTTP proxies, SOCKS proxies, and proxy credentials.
- Use included shell templates for forms, authenticated sessions, and content capture.

## Use Cases

- Validate Browser Flows: Run form, login, and navigation checks while collecting screenshots or video evidence.
- Capture Web Research: Open public pages, extract visible text, collect links, and save page screenshots for review.
- Automate Operations Tasks: Repeat authorized browser workflows in admin tools after confirming target sites and credentials.

## Prompt Templates

### Open a Page

```
Open this URL in a new browser session, show the visible interactive elements, and explain which element refs matter for the next step.
```

### Fill a Form

```
Navigate to this form, identify each required field, fill it with the provided test data, submit it, and verify the result page.
```

### Capture Evidence

```
Record the browser workflow, perform these steps, capture a full-page screenshot at the final state, and summarize what changed.
```

### Run an Authenticated Workflow

```
Use an authorized authenticated session to complete this workflow. Avoid recording sensitive fields, refresh refs after navigation, and close the session when done.
```

## Limitations

- Requires the inference.sh CLI and an authenticated CLI session before use.
- Element references can expire after navigation or dynamic page changes.
- Screenshots and videos are not automatically redacted for secrets or personal data.
- Browser actions can affect real accounts and sites, so authorization is required.

## Best Practices

- Use only authorized sites, accounts, and credentials for browser automation.
- Refresh snapshots after navigation, form submission, or dynamic page changes.
- Disable video recording when credentials, tokens, personal data, or internal records may appear.

## Anti Patterns

- Do not use proxy rotation to bypass rate limits, access controls, or site policies.
- Do not print cookies, passwords, tokens, or session data into logs or captured artifacts.
- Do not upload local files unless the target site, file path, and permission are explicit.

## Security Audit

- Audited at: 2026-07-06T14:21:40.441\+00:00
- Summary: Many static hits are false positives from Markdown code fences, relative documentation links, example URLs, and harmless /dev/null redirection. Confirmed risk remains significant because the skill runs an external browser automation CLI, supports screenshots, videos, file uploads, arbitrary page JavaScript, authenticated sessions, and SOCKS proxies. Additional semantic risks include cookie extraction guidance and proxy rotation for rate-limit avoidance.

## Stats

- Views: 97
- Downloads: 6
- Favorites: 0
- Popularity score: 0