Autonomous browser agent

Slapify

Give it a goal in plain English and it does what a person at a keyboard would: reads pages, clicks, types, logs in, fills forms, handles files and tabs, and keeps going until the job is done.

npx slapify init

View on npm GitHub Docs

0.0.23

Latest version

—

Weekly downloads

Modes

Slapify running an autonomous task in the terminal

Watch it work

Real runs, recorded by Slapify itself

Every run can be recorded: one annotated frame per step, a video and a replay you can scrub through. These two ran on demo sites.

Recorded Slapify run: Shops a store — **Shops a store**`Goal: Go to the store. Find a waterproof jacket that costs less than $150, add it to the cart in size M with gift wrap, and tell me the cart total.`Searches, compares, reads the product page, picks a size from a dropdown, ticks gift wrap and adds to cart.

Recorded Slapify run: Replies to posts — **Replies to posts**`Goal: Reply to the two posts about TypeScript with a useful tip each, and confirm each reply appears.`Finds the posts' real URLs with get_links, writes each reply, has it reviewed against a policy before it posts, then checks it appeared.

Capabilities

Does what a person at a keyboard does

Reads the page

get_page_text for posts, articles and results; get_links for real URLs, so it never guesses one.

Uses every control

Click, type, hover, double-click, drag, checkboxes, native dropdowns, even controls hidden for keyboard users.

Files and tabs

Uploads sandboxed to one folder, downloads with sanitized names, sign-in popups and new tabs followed.

Logs in

Credential profiles, cookie and session injection, and 2FA codes relayed over Telegram.

Sees when it has to

Optional vision for CAPTCHAs, charts and canvases. Screenshots go to the model as real images.

Stays undetected

Stealth mode patches only real automation tells, and persistent profiles keep you signed in.

Runs for days

Persistent runs only a human can stop, schedules, memory across sessions and hard time limits.

Records every run

Annotated frames, a video, a replay.html you can scrub through, and an optional live view.

Drives canvas apps

Held keys and coordinate clicks over CDP reach canvases and cross-origin iframes, games included.

Models

Runs on the plan you already pay for

Use a Claude or ChatGPT subscription instead of API credits, or any of six API providers with automatic fallback.

Claude subscription

provider: claude-cli

Runs every turn through your local claude -p on a Claude Pro or Max plan. No API key.

ChatGPT plan

slapify login codex

Any API

anthropic · openai · google · mistral · groq · ollama

List several and Slapify falls back down the chain when one rate-limits or goes down.

The harness

Autonomy you can leave alone

Tool contracts

Every tool declares what it can do to the world, from reading a page to buying something. Clicks are judged by the element they hit.

Admission, not loop guards

Repeats are never blocked. Repeated failures get guidance, and a post that already went through is held instead of sent twice.

Action review

Irreversible actions are checked against your plain-English policy first, and held when no reviewer can answer.

Recovery

Transient failures retry, outages and usage limits are waited out, and one bad step never ends a run.

Compaction

Long histories are summarised without ever paraphrasing the goal, so a week-long run keeps its instructions.

Fast decisions

Optional Jev or OpenAI Decisions answer "which element?" and "is this true?" in under a second, with the LLM as fallback.

How it works

Three modes, one engine

Task mode

Give it a goal in plain English. The agent plans, browses, logs in, recovers from errors and keeps going until the job is done, even if that takes days.

slapify task "…" for any goal
slapify fast for click-and-fill goals
Schedules, sessions and memory

Run mode

Executes .flow files: test scripts that describe user journeys step by step in plain English. Think Playwright, written like a checklist.

English-first .flow format
Popups handled, verify steps checked
CI/CD-friendly exit codes and reports

Programmatic mode

createAgent() runs the same harness as the CLI from JS or TS: inline config, streamed events and resumable sessions, plus perf audits and flows.

createAgent().run() and .stream()
Config inline or from .slapify/
Performance audits and flow runner

Quick start

Up and running in a minute

# Set up: pick an API key, or your Claude or ChatGPT plan
npx slapify init

# Give it a goal (and keep a replay of the run)
npx slapify task "Find a waterproof jacket under $150
  and add it to the cart in size M" --record runs/jacket

# Plain-English E2E tests
npx slapify run tests/checkout.flow --report

import { createAgent } from "slapify";

const agent = createAgent({ stealth: true });

const run = await agent.run(
  "Check the status of my last three orders on example-shop.com"
);
console.log(run.finalSummary);

for await (const e of agent.stream("Summarise my unread email")) {
  if (e.type === "status_update") console.log(e.message);
}