Agent docs

Two doors into the same engine: the MCP tools an agent calls directly, and the CLI the agent (or you) can drive from a terminal or script. Same daemon, same session, same knowledge cache either way. This page is the contract, not a tutorial.

MCP Tools

status

Full state in one call: API health, active session, stored credentials, and live browser windows.

browse

Headless fetch with session persistence. Fast, no interaction.

execute

Run a pipeline of browser steps. Auto-starts session, handles captchas, auto-escalates browser mode on bot-block.

session

Stop (save state) or clear (wipe) the browser session.

credentials

Store, list, or use encrypted login credentials. Store refuses to overwrite existing creds unless told to, blocks the 'login failed, re-ask for password' mistake.

knowledge

Query cached domain knowledge: login flows, selectors, known obstacles, captured API endpoints.

reverse-engineer

Capture a site's API: records the fetch/XHR calls made while driving it, auth, cookies, and saves .js/.ts client files.

remember

Save per-domain element anchors, target them as @a:<name> later, no re-finding. Self-healing: a failed anchor triggers re-discovery, not blind retries.

tabs

List and target tabs in your real Chrome (requires the browser extension).

clipboard

Read or write the machine's clipboard: read a code a site copied, stage text to paste.

Pipeline steps

Use with the execute tool. Steps run sequentially with a 20s stale-state timeout.

navigate

url

Go to URL

snapshot

List interactive elements with @e refs

find

role/name/text

Locate one element, return a ref

click

selector

Click element

human-click

selector | x,y

Human-like click

human-type

selector, value

Human-like typing

fill

selector, value

Fast fill (framework-aware events)

select

selector, value

Pick a native <select> option

read

selector

Read text content of an element

extract

expression

Evaluate JS, return result

evaluate

expression

Run JS (no return)

upload / paste / download

File upload, clipboard paste, download capture

screenshot

Capture current state

scroll

deltaY?

Scroll page

wait / wait-for

ms | selector

Wait for time or element

keyboard

key

Press key

domain

Fill login with stored creds

solve-captcha / recaptcha

Auto-detect and solve (requires ANTHROPIC_API_KEY)

API reverse engineering

Use the reverse-engineer tool to capture a site's API while your agent browses it. It records the fetch/XHR calls made while driving it, not static assets, including auth tokens, cookies, headers, and response bodies.

How to use it

"Reverse engineer the messaging API on this site"

The agent runs the same pipeline steps as execute, but records all network calls the page makes.

What it saves

auth.js

Shared cookies, tokens, Authorization header

getMessages.js

One file per endpoint: async function using fetch with auth

index.js

Re-exports all endpoint functions

README.md

Endpoint summary and dependency chain

Ask for TypeScript output to get .ts files with inferred types from the captured response bodies. A worked example: docs/example in the repo.

CLI reference

Everything above through a terminal instead of MCP, same daemon, same session. iframer-toolkit and iframer are the same binary; npx iframer-toolkit ... works without a global install.

iframer install

Install Chromium + MCP (skip if already installed)

iframer install mcp

iframer install extension chrome

Install the optional real-Chrome extension pairing host

iframer update [--check]

Update via npm (--check: report only)

iframer remove [chromium|mcp|extension]

Remove everything, or one piece

iframer status

Check system state and available modes

iframer execute <pipeline.json>

Run a pipeline, routes to the warm daemon (--json for raw)

iframer instances

List live browser windows (instanceId → current page)

iframer credentials add <domain>

Store encrypted login credentials

iframer credentials list / remove

List or delete stored credentials

iframer screenshot <url>

Take a screenshot

iframer browse <url>

Headless fetch with JS rendering

iframer reverse-engineer <url>

Capture a site's API to .js/.ts files

iframer session stop / clear

Save + close browsers, or wipe session data

iframer --cache / --cache <domain>

List cached domains, or print one domain's knowledge

iframer telemetry

Report estimated session tokens consumed by MCP tool calls

For the daemon, browser modes, and knowledge cache this all runs on, see architecture.

Home GitHub

Open source · MIT license · 2026