Why Browser Automation Fails and How OpenCLI’s API‑First Approach Solves It

Many enterprise systems run in browsers (admin consoles, ticketing platforms, deployment dashboards). Automating these interfaces can dramatically improve efficiency, but conventional UI‑driven agents often stumble on stability and speed.

Why Traditional Browser Automation Struggles

Agents that try to click buttons or fill forms must keep up with dynamic page changes, timing issues, and flaky selectors. Even small UI updates can break the automation, making it unreliable for production workloads.

OpenCLI’s Core Idea

Instead of fighting the UI, OpenCLI captures the underlying API calls that the front‑end uses. By reverse‑engineering these endpoints and reproducing the requests directly, scripts become far more robust and faster.

Quick Start

npm install -g @jackwener/opencli

Basic commands:

opencli list                     # list all commands
opencli list -f yaml             # list in YAML format
opencli hackernews top --limit 5   # public API, no browser needed
opencli bilibili hot --limit 5      # browser‑based command
opencli zhihu hot -f json           # JSON output
opencli zhihu hot -f yaml           # YAML output

AI Agent Workflow for API Discovery

Open Browser – navigate to the target page.

Observe Page – snapshot interactive elements (buttons, links).

First Capture – record network requests, filter JSON API endpoints, note URL patterns.

Simulate Interaction – click elements (e.g., subtitles, comments, follow) using browser_click and wait for new requests.

Second Capture – compare with the first capture to identify newly triggered APIs.

Validate API – fetch the endpoint with proper credentials and verify the response structure.

Write Adapter – generate TypeScript or YAML adapters based on the confirmed API.

Five‑Tier Authentication Strategy

OpenCLI classifies APIs into five tiers based on the required authentication:

Tier 1 – Public : Direct fetch(url) works without cookies.

Tier 2 – Cookie : Requires session cookies; use fetch(url, {credentials:'include'}).

Tier 3 – Header : Needs additional headers (e.g., Bearer token, CSRF).

Tier 4 – Intercept : Data is hidden in client‑side stores (Pinia/Vuex); capture via store actions or XHR interception.

Tier 5 – UI : Only reachable through UI automation; last resort.

Adapter Generation

Adapters can be written in TypeScript for complex scenarios or YAML for simple ones. Files are stored under ~/.opencli/clis/{site}/{command}.ts|yaml. Example workflow:

# Create directory
mkdir -p ~/.opencli/clis/{site}
# Generate adapter (YAML or TS)
# Verify
opencli list | grep {site}
opencli {site} {command} {option}

External CLI Integration

Existing CLI tools can be wrapped by OpenCLI, allowing seamless reuse of established commands within the OpenCLI ecosystem.

AI‑Native CLI Synthesis

Explore & analyze: deep page crawling, auto‑scroll, network interception, framework detection.

Strategy selection: automatically choose the appropriate authentication tier.

Adapter synthesis: generate candidate YAML/TS files with templated URLs, field mappings, and default parameters.

Test & validate: run the generated pipeline, fallback if needed.

Record Mode

The opencli record command records user interactions in the browser, scores and semantically analyzes the captured request sequence, and produces reusable CLI commands.

Current Limitations

Payload capture is incomplete: only request metadata (URL, method, response body) is stored; POST/PUT bodies are missing.

Generation is limited to read‑only APIs; write operations (create, update, delete) cannot yet be auto‑generated.

Future Outlook

Software competition will shift from UI polish to callable APIs. Agents care only about stable, well‑defined commands, parameters, and responses. Making software easily discoverable and executable by AI agents will become a decisive advantage.