AI Agents

Agent integration flow

se-cli works with any AI agent that can run shell commands — Claude Code, Cursor, Copilot CLI, and more. There is no MCP server required for the CLI path: install once, drop in SKILL.md, and the agent drives the browser like a human would.

1. Install the CLI

One global install — no per-agent setup or MCP server configuration needed.

npm install -g @browsers-cli/se-cli

2. Install the skill (recommended)

install --skills copies SKILL.md into your agent's skills directory so it learns the command syntax instantly. Targets Claude Code (.claude/skills/se-cli/), Cursor (.cursor/skills/se-cli/), GitHub Copilot (.github/copilot/skills/se-cli/), and a generic .agents/skills/ path (v0.9).

se-cli install --skills                  # auto-detect installed agent dirs
se-cli install --agent=claude,cursor     # explicit multi-target
se-cli install --list-agents             # show supported targets
se-cli install --force                   # overwrite existing files

3. Agent drives the browser

The agent calls se-cli commands like a human would — open, snapshot, interact, close.

User: "Check that the login page works"
Agent: I'll navigate and test the login flow.
  $ se-cli open https://app.example.com/login
  $ se-cli snapshot
  $ se-cli fill e1 "test@example.com"
  $ se-cli fill e2 "password"
  $ se-cli click e3
  $ se-cli snapshot
  Login succeeded — dashboard is visible.

4. Why CLI beats MCP for coding agents

No tool schemas pollute the context window. No JSON-RPC envelope overhead. Just clean shell commands and compact YAML output — typically ~12× fewer tokens than Selenium MCP per interaction.

MCP vs CLI token cost

For a coding agent that already runs shell commands, the CLI path is dramatically cheaper. The MCP Server (below) is for autonomous workflows that need a persistent tool server instead of a shell.

Aspect	Selenium MCP	se-cli (CLI)
Tokens / call	~5,000	~400
Schema in context	~5KB, always loaded	None
Output format	JSON-RPC envelope + full tree	Compact YAML + refs
Agent requirement	MCP client	Any shell-capable agent
Persistent server	Yes	No (daemon holds state)
Best for	Autonomous long-running workflows	Coding agents (Claude Code, Cursor)

MCP Server (Shipped · v0.9)

se-cli ships as an MCP Server (custom JSON-RPC 2.0, no SDK dependency), dual-tracking the CLI: both share the same underlying tool implementation. This covers autonomous agents that prefer a persistent tool server over shell commands.

# stdio transport (default) — local agent
se-cli mcp-server

# thin npm wrapper (recommended for MCP clients)
npx @browsers-cli/se-mcp

• All CLI tools wrapped as registerTool calls
• run-code "async driver => ..." for arbitrary Selenium snippets
• generate-locator <ref> for best-available locators
• Role-based codegen (By.role()) via the shared codegen module
• SKILL.md frontmatter compliance (name, description, license, compatibility)

Dual-track strategy

CLI + SKILLS stays the token-efficient path for coding agents; the MCP Server adds persistent-state tooling for autonomous workflows. Track progress on the roadmap.

v0.10: remote, custom browsers & Safari (Shipped)

For agents that need browsers beyond the local Chrome/Edge/Firefox defaults:

# Attach to a Selenium Grid 4 hub or remote WebDriver
Agent: "Run the test against the staging Grid"
  $ se-cli open --browser=chrome --endpoint=http://grid.example.com:4444/wd/hub
  $ se-cli grid status http://grid.example.com:4444/wd/hub
  $ se-cli grid distribute --shard=1/4 --browsers=chrome,edge,firefox,safari

# Custom browser/driver binaries (360, UC, Brave, Electron-embedded, ...)
  $ se-cli open --browser-binary=/opt/custom-browser --driver-binary=/opt/custom-driver

# Real Safari (macOS only — no headless, no BiDi/CDP)
  $ se-cli open --browser=safari

# W3C capabilities pass-through (e.g. accept self-signed certs)
  $ se-cli open --capabilities='{"acceptInsecureCerts":true}'

Not every command works on every browser. Navigation, snapshot, interaction, cookies, storage, screenshot, eval, tabs and dialogs work on all four; console/requests/route need BiDi (Chrome/Edge/Firefox), device/ emulate/--cdp need CDP (Chrome/Edge), pdf needs the W3C print endpoint (no Safari). The full matrix lives in the se-cli repo docs (docs/spec.md → Browser Support Matrix). Agents that try an unsupported command get a clear error naming the missing capability, so they can fall back to an alternate browser automatically.