Reference

Use live help as the contract for flags and commands.

The docs summarize stable command families. The exact parser surface is the output of agbrowse --help and agbrowse web-ai --help.

Skill commands

skills, skills list, skills get, skills path, skills install. Query and install bundled agent skills (browser, web-ai, search, vision-click). Use --target to choose a skill root, --link to track the global npm package, and --force to replace existing skills.

Browser lifecycle

start, stop, status, reset. Manage the Chrome CDP connection. start accepts --port, --headless, and --chrome-path.

Observe

snapshot, screenshot, text, get-dom. Capture DOM refs, page text, screenshots, or raw DOM fragments.

Act

click, type, press, hover, select, check, uncheck, drag, mouse-click, move-mouse, mouse-down, mouse-up. Ref-based or coordinate-based page mutations.

Navigate

navigate, reload, resize, scroll, tabs, active-tab, tab-switch, select-tab, new-tab, tab-close, tab-cleanup. Navigation, viewport, tab management, and active target resolution.

Wait

wait, wait-for-selector, wait-for-text, wait-for. Time-based, selector, text, or legacy ref waits.

Diagnostics

console, network, evaluate. Capture console/network logs or run arbitrary page JavaScript.

URL read

fetch <url> — adaptive 6-phase escalation ladder for reading a single candidate URL. Key flags: --json, --trace, --browser auto|never|required, --browser-session none|isolated|existing|user|interactive, --identity auto|minimal|chrome, --allow-third-party-reader, --max-bytes N, --timeout-ms N, --selector CSS.

Search commands

search "<query>" — standalone deep search: query rewrite, candidate discovery via adaptive-fetch ladder (TLS-impersonation, Camoufox, feed-parser, BM25, table-extractor, candidate-discovery), and evidence scoring. Flags: --json, --deep, --verify <url>, --stdin-results, --browser auto|never, --max-results N, --vendor grok|chatgpt|gemini.

Research commands

research plan, research normalize-results, research enrich-fetch, research browse-plan. Offline planning surface for Korean/source-sensitive search: decompose queries into constraints and candidate URLs, normalize provider search results, enrich URLs with adaptive fetch, and plan browser escalation. None of these commands run live searches or mutate browser state.

Runway commands

runway selectors, runway status, runway open, runway preflight, runway poll. Read-only preflight for Runway Apps/Custom media task-runner. Safety contract: generate, submit, and payment controls are never clicked.

Web-AI commands

render, status, send, poll, query, code, code-extract, stop, watch, snapshot, observe-bundle, observe-actions, upload, sessions list|show|resume|reattach|doctor|prune, project-sources list|add, context-dry-run, context-render, doctor, claim-audit, eval, mcp-server, work send.

ChatGPT Deep Research

--research deep activates ChatGPT Deep Research mode (experimental). Saves a report artifact, records researchMode: "deep" in the session, skips auto archive, and auto-confirms the post-submit plan card. ChatGPT-only, cannot be combined with --follow-up.

ChatGPT image output

--output-image ./out.png saves ChatGPT-generated images. Multiple images are written as sibling files (out.png, out-2.png, out-3.png). Fail-closed: if no image can be detected or saved, the command returns provider.image-output.

ChatGPT follow-ups

--follow-up "..." appends sequential follow-up prompts in the same command. ChatGPT-only. Cannot be combined with --research deep.

Help checks

agbrowse --help
agbrowse web-ai --help
agbrowse web-ai code --help
agbrowse web-ai code-extract --help
agbrowse web-ai work --help
agbrowse runway --help
agbrowse search --help
agbrowse research --help
agbrowse skills --help

ChatGPT Work

agbrowse web-ai work send --prompt "Analyze this repository" --power 4

--power accepts 1 through 6. MCP clients use web_ai_work_send. Chat commands and web_ai_submit_prompt reject Work; they do not accept surface=work.

ChatGPT model inputs

Legacy effort remapping: light|low|standard|normal|regular|default → medium, extended → high, heavy|extra-high|extra_high|extra high → xhigh. Extended emits exactly one stderr warning after selecting High. Legacy Pro effort resolves to flat Pro and emits the same one-line no-selection warning pattern.

Timeout resolution

Priority: explicit timeout, stored deadline remainder, tier default, vendor fallback. Unknown tiers fall back to ChatGPT 1200, Gemini 1200, or Grok 600 seconds.

The user-reported roughly 40-minute ChatGPT Pro budget was not present in the inspected DOM and is not the agbrowse timeout.