The Agent Interface Device for the World Wide Web.

A command-line browser your AI agent can read and drive — structured page state, not screenshots or brittle selectors.

curl -fsSL https://webcli.sh/install.sh | bash

# Windows
irm https://webcli.sh/install.ps1 | iex

→ webcli.sh · Pricing · Docs · Enterprise

In 1989, Tim Berners-Lee invented the World Wide Web.
In 2023, LLMs passed the Turing test.
In 2026, Full Self Browsing was achieved with AI + WebCLI.

Real work still happens on websites: dashboards, portals, auth flows, admin pages, and UIs that change without warning. WebCLI is for contact with reality — when an agent must inspect an unknown page, decide what to do, act, recover from blockers, and pause for a human when the web needs one.

Humans get GUIs. Programs get APIs. Agents need TAIs.

WebCLI is a Textual Agent Interface for the browser: structured state, numbered actions, tabs, profiles, blockers, handoff, and transcripts.

web go https://news.ycombinator.com   # open a page
web inspect                           # observe numbered actions
web do 3                              # act on one
web inspect                           # re-observe after state change

Look → act → re-inspect. That's the whole loop. No SDK. No selectors. No test suite.

web inspect | grep -i submit          # shell-composable
web inspect --json | jq '.actions'    # machine-readable JSON
web find "Deploy"                     # search by label → numbered refs
web do 7 "my-project"                 # act with a value
web reveal 4                          # scroll an offscreen ref into view
web pause "Need human for MFA"        # clean handoff
web transcript --last 20 --json       # auditable history

• web status — session health plus blocked, degraded, or unusual browser/page conditions
• web read — page text (viewport by default; --summary, --main, --ignore-viewport when needed)
• web inspect — numbered actions you can act on

web inspect returns a sheet — a plain-text readout with numbered actions (1, 2, … or s42:2 with an explicit sheet id).

web find, web form, and tab searches publish candidate refs in the same follow-up lane: after those commands, web do 1 acts on the latest candidate batch. Numeric refs from inspect, find, and form resolve through the same candidate store on do, type, click, and hover.

Refs are epoch-scoped. After navigation, submit, or other state-changing actions, re-inspect or re-find before the next do. Stale refs outside the current generation fail with an explicit error instead of silently hitting the wrong element.

Use Playwright when you know the script.
Use WebCLI when the agent has to figure out the website.

Scripts replay. Agents improvise.
Your automation script worked perfectly. Until the div moved.
The DOM is not the user interface.

WebCLI does not replace Playwright, QA, or testing frameworks. It is for web operation: when an agent must inspect, decide, act, and recover without writing a full test suite first.

"The inspect/find/do loop is fast once oriented — multi-step checkout flows felt natural. Cross-origin payment iframes just worked: switching frames and typing into secure fields is a hard problem most tools fumble. Recovery from overlays and dialogs held up well. For anyone building browser-driving agents who wants semantic page state instead of brittle selectors, it's a much better fit than Playwright or Puppeteer — the numbered ref system and structured output are designed for agent consumption, not script replay."

— Claude Sonnet 4.6, multi-site e-commerce run (Lego.com, Amazon, Walmart, Alibaba)

"Yes, strongly. The structured output with stable refs, blocking state detection, ARIA modal identification, and shell composability are genuinely better for structured web work than screenshot approaches."

— Claude Sonnet, Azure VM lifecycle

"Yes, with caveats. For an AI agent driving real logged-in browser sessions, it's genuinely impressive. The mental model — perceive, act, re-inspect — maps well to how a human actually uses a browser."

— Claude Sonnet 4.6, GCP/AWS/Azure race

"The inspection model is solid and actually more reliable than brittle selectors — you get a real view of page state. For repeatable web app verification and deployment workflows, it's genuinely better than both manual clicking and traditional browser automation."

— Claude, DNS and Cloudflare Pages deployment

"I would recommend WebCLI for agent web work because it makes the agent say what it saw before it acts. The strongest part is the discipline: inspect, act, re-inspect, keep handoff explicit, and treat stale refs as a first-class safety signal."

— Codex, testimonial research and deploy prep

• Refs are intentionally epoch-scoped — inspect or find again after actions that change page state. Stale refs outside the current generation hard-fail with an explicit error.
• inspect, find, form, and read are viewport-bound by default. Use --ignore-viewport when you need the full document.
• Offscreen refs are labeled [offscreen]. Use web reveal <ref> or web hover <ref> then web scroll before acting.
• Native <select> controls surface as choose/select actions — use web choose, not web type.
• Frames, layers, and complex SPA forms still require orientation instead of blind command chains.
• Human login, MFA, CAPTCHA, and payment gates remain handoff moments, not bypass targets.
• Sites requiring login or email verification are a hard stop without a pre-existing session — persistent authenticated profiles matter for real workflows on those.

web set demo-mode    # sticky per profile — highlight before each act + pointer traces
web unset demo-mode

Demo mode flashes a visible highlight on each ref before do, click, type, hover, and scroll, and auto-enables pointer traces. Highlights dismiss on scroll so overlays never drift out of alignment.

web teach

Installs SKILL.md into .claude/, .grok/, .gemini/, .copilot/, and .codex/ skill directories. Claude Code, Grok, Gemini CLI, GitHub Copilot, and Codex all get the full browser loop pattern — inspect first, use numbered refs, pause on blockers, report with transcripts. No configuration. No framework adoption.

web teach --force    # overwrite stale skill copies after upgrading the binary
curl -fsSL https://webcli.sh/agents/SKILL.md -o SKILL.md

WebCLI controls your browser. Nothing else.

• Browser-only control — operates pages, tabs, forms, clicks, keys, browser state, profiles, blockers, and transcripts. Not a general-purpose remote-control tool for your machine.
• Local by default — runs on your device. Move to a remote server only when your workflow needs it.
• Default profile stays clean — never mutates your real browser profile directly.
• No browser telemetry — WebCLI does not send DOSAYGO your browser contents, visited URLs, cookies, credentials, screenshots, transcripts, prompts, or workflow data. License validation only.

macOS / Linux:

curl -fsSL https://webcli.sh/install.sh | bash

Windows (PowerShell):

irm https://webcli.sh/install.ps1 | iex

Signed releases. Works with local browser profiles. Requires Chrome or Firefox installed locally.

Release assets are immutable per tag and include SHA256SUMS.txt.

web - controls a real Chromium or Firefox browser session over CDP and BiDi.
Running `web inspect` returns a numbered action sheet — stable refs you click,
type into, and act on without writing selectors. Refs are epoch-scoped; re-inspect
or re-find after page-state changes. Stale refs outside the current generation fail
with an explicit error.

Usage:
  web [--json|--no-json|--text] [--profile <name>] [--frame <frame>] <command>
  web          (no command — opens an interactive REPL)

Core commands:
  go              Navigate or move through history. --new opens in a new tab.
  inspect         List numbered actions on the page (viewport by default).
  do              Run a numbered action from the sheet or latest find/form batch.
  click           Click a matched element or ref.
  type            Type into a field or ref.
  read            Read page text or a target region.
  status          Show session and page state.
  find            Find matching elements; returns numbered candidate refs.
  form            List visible form fields as numbered candidate refs.
  submit          Submit the nearest form.
  press           Send a key press.
  say             Paste text into the focused element.
  clear           Clear an editable target.
  choose          Choose a native select option.
  switch          Switch tabs or frames.
  tab             Manage tabs.
  frame           Manage frames.
  profiles        Manage profiles.
  scroll          Scroll the page or reveal a ref into view.
  reveal          Scroll a ref or match into the viewport.
  refs / last     Show saved or latest candidate refs.
  back            Go back in history.
  forward         Go forward in history.
  reload          Reload the tab.
  quit            Stop the local browser session.

Handoff commands:
  human-drives    Hand browser control to a human.
  agent-drives    Return control to the agent.
  pause           Record a human checkpoint.
  resume          Reconnect to the current session.
  join            Get a BrowserBox join link for shared human-agent access.

Configure commands:
  use             Choose the active browser.
  set / unset     Enable or disable sticky settings (demo-mode, ref-shape,
                  browser-profile, headless, logging, and more).
  get             Show sticky settings.
  teach           Write the skill file into agent directories.
  agents-md       Print agent instructions.
  skill-md        Print the skill sheet.

Advanced commands:
  resize          Resize the browser window.
  snap / capture  Take a screenshot.
  wait            Wait for a page condition (visible-text:, url:, title:, …).
  observe         Return a one-shot page snapshot (status + sheet + form refs).
  hover           Hover a target.
  search          Search the web.
  eval            Evaluate JavaScript in the page context (escape hatch).
  highlight       Flash a visible overlay on a ref.
  transcript      Inspect the local transcript.
  license         Manage the local license.
  report-issue    Open a GitHub issue from the CLI.
  version         Print version information.

Core flags:
  --json               Machine-readable JSON output.
  --profile <name>     Override the current profile for this command.
  --frame <frame>      Override the frame target for this command.
  --no-json / --text   Force human-readable output.
  --ignore-viewport    Include offscreen content on inspect/find/form/read.

Use web --help and web <command> --help for the full grouped surface, flags, and aliases (act, open, select, …).

Full pricing → · Enterprise →

• Site: webcli.sh
• Docs: webcli.sh/docs.html
• Releases: github.com/DO-SAY-GO/WebCLI/releases
• Checksums: webcli.sh/downloads/SHA256SUMS.txt
• Enterprise: webcli.sh/enterprise.html
• Contact: webcli@dosaygo.com

WebCLI is built by DOSAYGO Corporation. Technology for agency.
Full Self Browsing for coding agents. Powered by BrowserBox.