# Prompteryx

> Prompteryx is an AI platform where one chat does the work: it browses real websites for you, runs automations, researches with sources and makes websites, documents, images, video and music.

Prompteryx is an AI workspace for delegating real work. A single chat picks the right AI model for each job, uses an AI browser agent that operates a real Chrome browser (your own through Prompteryx Desktop, or a cloud browser), connects to more than 3,000 apps, and can turn any task into a repeatable workflow on the Visual Studio canvas that runs on a schedule, by API or by webhook.

Prompteryx is made by PROMPTERYX LIMITED, a company registered in England and Wales (company number 17472473).

## Key facts
- One chat for everything, with Auto model selection across leading AI models
- AI Browser Agent that uses a real Chrome browser to complete tasks on any website
- Prompteryx Desktop (Windows) to run agents in your own Chrome with your own logins
- Cloud browsers with residential proxies, session recording and automatic CAPTCHA solving
- Visual Studio: a drag-and-drop workflow canvas with 80+ browser action nodes
- Connect Hub: actions in 3,000+ apps such as Gmail, Google Sheets, Slack and Notion
- Schedules, webhooks and a REST API to run any workflow automatically
- Batch Prompter for running many prompts across AI platforms from a spreadsheet
- Websites, documents, slide decks, images, video, voice and music made in the chat
- Explore: a gallery of things people made with Prompteryx, with the recipe behind each one
- Developer tools: REST API, TypeScript SDK, CLI and a hosted MCP server that Claude, ChatGPT, Cursor and VS Code connect to with OAuth; n8n, Make and Zapier connect through their HTTP, webhook and MCP apps
- Pricing: A 7-day free trial, then monthly plans (Starter, Developer and Pro, with Enterprise on request). Each plan includes AI Credits, workflow executions and cloud-browser time; extra usage can be paid as you go from AI Credits. Current prices: https://www.prompteryx.ai/pricing

## Key pages
- [Home](https://www.prompteryx.ai/)
- [Pricing](https://www.prompteryx.ai/pricing)
- [Why Prompteryx](https://www.prompteryx.ai/why-prompteryx)
- [Templates and examples](https://www.prompteryx.ai/templates)
- [Use cases](https://www.prompteryx.ai/use-cases)
- [Integrations](https://www.prompteryx.ai/integrations)
- [Documentation](https://www.prompteryx.ai/docs)
- [Download Prompteryx Desktop](https://www.prompteryx.ai/download)

## Products
- [The Prompteryx app (one chat for everything)](https://www.prompteryx.ai/features/app)
- [AI Browser Agent](https://www.prompteryx.ai/features/ai-browser-agent)
- [Visual Studio workflow builder](https://www.prompteryx.ai/features/visual-studio)
- [Cloud Browser](https://www.prompteryx.ai/features/cloud-browser)
- [Batch Prompter](https://www.prompteryx.ai/features/batch-prompter)

## Comparisons
- [Compare Prompteryx](https://www.prompteryx.ai/compare)
- [Prompteryx vs Browserbase](https://www.prompteryx.ai/compare/vs-browserbase): Prompteryx and Browserbase both run cloud browsers for automation and AI agents. Compare browser-hours, billing, regions, SDKs, no-code building, your own Chrome and app connections. Facts checked 25 October 2026.
- [Prompteryx vs n8n](https://www.prompteryx.ai/compare/vs-n8n): n8n is a popular workflow automation platform you can self-host. Prompteryx adds a built-in browser engine, AI browser agents, recordings of every run and your own Chrome. A fair comparison, checked 25 October 2026.
- [Prompteryx vs Zapier](https://www.prompteryx.ai/compare/vs-zapier): Zapier connects 9,000+ apps through their APIs. Prompteryx automates the steps that have no API: websites, portals and forms, with an AI browser agent. A fair comparison, checked 25 October 2026.
- [Prompteryx vs Manus](https://www.prompteryx.ai/compare/vs-manus): Manus runs general-purpose AI agents in the cloud and on your computer. Prompteryx combines AI browser agents with workflows you can edit step by step and replay without AI costs. A fair comparison, checked 25 October 2026.
- [Prompteryx vs Claude](https://www.prompteryx.ai/compare/vs-claude): Claude is a leading AI assistant with Claude in Chrome, computer use, connectors and scheduled tasks. Prompteryx is an automation platform that can use Claude models. Compare browser agents, workflows, schedules and cloud browsers. Checked 25 October 2026.
- [Prompteryx vs Perplexity Comet](https://www.prompteryx.ai/compare/vs-comet): Comet is Perplexity's free AI web browser with a built-in assistant. Prompteryx is an automation platform with AI browser agents, cloud browsers and workflows that replay without AI costs. A fair comparison, checked 25 October 2026.
- [Prompteryx vs Axiom.ai](https://www.prompteryx.ai/compare/vs-axiom): Axiom.ai builds no-code browser bots in a Chrome extension, with cloud runs and a desktop app. Prompteryx adds an AI browser agent, a drag-and-drop canvas and 3,000+ app actions. A fair comparison, checked 25 October 2026.

## Company, trust and support
- [Enterprise](https://www.prompteryx.ai/enterprise)
- [Trust centre](https://www.prompteryx.ai/trust)
- [Subprocessors](https://www.prompteryx.ai/trust/subprocessors)
- [Security](https://www.prompteryx.ai/security)
- [Support](https://www.prompteryx.ai/support)

## Documentation
- [AI Browser Agent documentation](https://www.prompteryx.ai/docs/ai-browser): An AI browsing copilot with computer vision that runs on our Cloud Browser or in your own Chrome
- [AI Browser Agent in the chat](https://www.prompteryx.ai/docs/app/browser-agent): Run the AI Browser Agent from the main chat: choose the cloud browser or your own Chrome, pick a model, watch it work, and read the outcome card with Try again and Try with another model.
- [The app: Connectors](https://www.prompteryx.ai/docs/app/connectors): Connect the apps you use, such as Gmail, Google Sheets, Slack and Notion, so the chat can read and act in them for you.
- [What you can make](https://www.prompteryx.ai/docs/app/create): One chat makes websites, games, images, video, music, voiceovers, sound effects, documents, slides, spreadsheets, PDFs and cited research. How to ask, what you get back, and how to keep, publish or share it.
- [Prompteryx Desktop](https://www.prompteryx.ai/docs/app/desktop): Prompteryx Desktop runs automations in your own Chrome with your logins and free local minutes, and opens the Prompteryx app in its own window.
- [Explore and publishing](https://www.prompteryx.ai/docs/app/explore): Explore is where people share what they make with Prompteryx. Browse websites, games, images, videos, music and browser tasks, see the prompt behind each one, make your own version, and share your own work.
- [The app: Groups](https://www.prompteryx.ai/docs/app/groups): Group chats and direct messages where anyone can ask @Prompteryx to make something for everyone.
- [Memory and Skills](https://www.prompteryx.ai/docs/app/memory-and-skills): What the Prompteryx assistant remembers about you and the switches that control it, the About you profile, incognito chats, and reusable Skills.
- [Models and Auto](https://www.prompteryx.ai/docs/app/models): How the Prompteryx app chooses a model with Auto, how to pick a specific model for one message, and how to change your defaults.
- [The Prompteryx app: documentation](https://www.prompteryx.ai/docs/app): The Prompteryx app is where everything starts: one chat that answers with any model, builds websites and games, makes images, video, music and documents, does research, runs the AI Browser Agent and builds workflows.
- [Your personal agent (beta)](https://www.prompteryx.ai/docs/app/personal-agent): An agent with its own name that keeps working when you're not in the app: it briefs you, watches pages for you, and asks before it acts.
- [The app: Phone app](https://www.prompteryx.ai/docs/app/phone): Prompteryx on Android: your chats, Groups, Studio and your agent, with notifications.
- [Plans, credits and top-ups](https://www.prompteryx.ai/docs/app/plans-and-credits): How Prompteryx is paid for: the free trial, monthly plans, AI credits, one-off top-up packs, pay-as-you-go usage credits, and changing or cancelling your plan.
- [Projects in the Prompteryx app](https://www.prompteryx.ai/docs/app/projects): Projects in the Prompteryx app: group chats around a goal, with instructions every chat follows, files every chat can use, and team members.
- [Schedule tasks from the chat](https://www.prompteryx.ai/docs/app/schedules): Run any chat task on a schedule from the message box: how often, where it runs, what happens when a run needs you, and where your schedules are listed.
- [Welcome, settings and skins](https://www.prompteryx.ai/docs/app/settings): The welcome set-up for new accounts, what is in Settings, choosing light or dark and a skin, and where your account, plan and usage live.
- [Share links for chats](https://www.prompteryx.ai/docs/app/sharing): Share a read-only copy of a Prompteryx chat with a link: what is included, what is hidden, how to update or remove a link, and publishing to Explore.
- [Batch Prompter](https://www.prompteryx.ai/docs/batch-prompter): Run batch automations across AI platforms with Excel files
- [Migrate to Prompteryx Cloud Browser](https://www.prompteryx.ai/docs/cloud-browser-migration): Drop-in migration from Browserbase, Browser Use Cloud, or Puppeteer Cloud to the Prompteryx Cloud Browser — same CDP connect, lower cost.
- [Cloud Browser quickstart](https://www.prompteryx.ai/docs/cloud-browser-quickstart): Spin up a real Chrome browser in the Prompteryx cloud and drive it with Playwright in under five minutes. API keys, SDK, and Sessions API copy-paste ready.
- [Prompteryx Cloud Browser: overview and API](https://www.prompteryx.ai/docs/cloud-browser): Real Chrome browsers in the Prompteryx cloud, driven from your code over CDP: recordable, proxy-capable and metered to your plan. Or skip the session with managed Fetch and Search.
- [Run the AI browser agent](https://www.prompteryx.ai/docs/developers/agent): Describe a goal. Prompteryx opens a browser, looks at the screen, clicks and types until the goal is met, and reports what it found and what it cost.
- [Authentication & API Keys](https://www.prompteryx.ai/docs/developers/authentication): Create and use API keys to authenticate with all Prompteryx developer interfaces
- [Drive a cloud browser](https://www.prompteryx.ai/docs/developers/browser): Start a real Chrome in the cloud, connect your own Playwright to it, and borrow AI only for the steps that are hard to script.
- [Prompteryx CLI](https://www.prompteryx.ai/docs/developers/cli): Prompteryx from a terminal or a CI job. Sign in once in the browser; tokens refresh on their own.
- [Use connected apps (Connect Hub)](https://www.prompteryx.ai/docs/developers/connect-hub): Send an email, add a row, post a message: 3,000 apps through the user's own connected accounts, with credentials that never leave the platform.
- [Prompteryx for Make.com](https://www.prompteryx.ai/docs/developers/make): Run Prompteryx workflows and the AI browser agent from Make scenarios with the HTTP app, receive results with a custom webhook, or call Prompteryx tools through Make's MCP Client
- [Prompteryx MCP server](https://www.prompteryx.ai/docs/developers/mcp): Use Prompteryx workflows, the AI browser agent and your connected apps from Claude, ChatGPT, Cursor, VS Code, Gemini CLI and any other MCP client
- [MCP Tools Reference](https://www.prompteryx.ai/docs/developers/mcp/tools): Complete reference for all browser, workflow, integration, AI, and data tools available via MCP
- [Prompteryx for n8n](https://www.prompteryx.ai/docs/developers/n8n): Run Prompteryx workflows and the AI browser agent from n8n, get results back by webhook, or give an n8n AI Agent the Prompteryx tools over MCP
- [Prompteryx developer documentation](https://www.prompteryx.ai/docs/developers): Describe a goal in plain English. Prompteryx opens a browser, does the task, and can save the path as a reusable workflow.
- [REST API Reference](https://www.prompteryx.ai/docs/developers/rest-api): Programmatically execute and manage your automation workflows via REST API
- [Prompteryx TypeScript SDK](https://www.prompteryx.ai/docs/developers/sdk): The REST API with the typing done for you. Node 18+, ESM and CommonJS, browsers and edge runtimes with a global fetch.
- [Webhooks](https://www.prompteryx.ai/docs/developers/webhooks): Trigger workflows from external events, and receive a signed callback when a run finishes
- [Run workflows from anywhere](https://www.prompteryx.ai/docs/developers/workflows): Anything built on the Visual Studio canvas, or saved from an agent run, has an id. Run it with variables, wait for the result, put it on a schedule, or trigger it from a webhook.
- [Prompteryx for Zapier](https://www.prompteryx.ai/docs/developers/zapier): Start Prompteryx workflows from any Zap and get the result back in another Zap, with Webhooks by Zapier
- [Workflow node reference](https://www.prompteryx.ai/docs/nodes): Every building block you can drop onto the canvas, organized by category
- [Documentation](https://www.prompteryx.ai/docs): Documentation for Prompteryx: the app, the AI Browser Agent, Visual Studio workflows, Prompteryx Desktop, cloud browsers, integrations and the API.
- [Visual Studio workflow builder documentation](https://www.prompteryx.ai/docs/visual-builder): Drag-and-drop automation builder featuring 70+ browser nodes and 200+ integrations

## Optional
- [Everything in one file (all docs, plans, Q&A, use cases)](https://www.prompteryx.ai/llms-full.txt)
- [Claude skill for the API](https://www.prompteryx.ai/SKILL.md)
- [Privacy policy](https://www.prompteryx.ai/privacy)
- [Terms of service](https://www.prompteryx.ai/terms)

---

# REST API guide for AI coding agents
For coding agents (Claude Code, Cursor, ChatGPT…) helping a user drive Prompteryx. Generated from the live
route code and verified against the deployed API; prefer it over /api/v1/openapi.json (partial) and SDK docs.

## What the API does
Prompteryx is a browser-automation platform: AI agents that drive a real Chrome in the
cloud, a visual workflow builder, and a REST API. Hand it a natural-language goal and get
back the answer, a full step trace, and exact usage/cost telemetry. Base URL: https://www.prompteryx.ai

## Auth — TWO separate key families (do not mix them up)
1. Platform key  "px_live_..."   -> header  Authorization: Bearer px_live_...
   Used by /api/v1/* (AI agent, workflows, executions, schedules, subscription, ...).
2. Cloud-browser key "pcb_live_..." -> header  x-api-key: pcb_live_...
   REQUIRED for /api/v1/cloud-browser/* (sessions, fetch, search). A px_live_ key is
   rejected with 401 on this group.
Both are created signed-in at https://www.prompteryx.ai/cloud-platform/api?tab=keys
("Platform API keys" card and "Cloud-browser keys" card). Each key is shown ONCE.

## Limits and billing
- Rate limit per platform key: 60 requests/minute, 10,000/day. Responses carry
  X-RateLimit-Limit-Minute / -Remaining-Minute / -Limit-Day / -Remaining-Day; over the
  limit -> 429 RATE_LIMIT_EXCEEDED with Retry-After (seconds).
- Billing unit = AI Credit (1 credit is roughly $0.01 of model cost). Runs also meter
  cloud-browser minutes / proxy MB / CAPTCHA solves against the account's plan.
- ALWAYS set "maxCredits" on agent runs — it is a hard cost cap; the run stops there.

## Response envelope (platform /api/v1/* routes)
Success: { "success": true, "data": { ... }, "meta": { "requestId", "timestamp" } }
Error:   { "success": false, "error": { "code", "message" } } + proper HTTP status.
Common codes: MISSING_API_KEY / INVALID_API_KEY / API_KEY_EXPIRED (401), UNSUPPORTED_MODEL (400),
QUOTA_EXHAUSTED (402 = out of AI credits; 429 = monthly workflow runs used up),
QUOTA_EXCEEDED (Connect Hub calls: 429 + Retry-After, or 402 = not on the plan),
FORBIDDEN / INSUFFICIENT_SCOPE (403), NOT_FOUND (404), RATE_LIMIT_EXCEEDED (429 + Retry-After).
Do not retry QUOTA_* immediately; only RATE_LIMIT_EXCEEDED clears on its own.
The /api/v1/cloud-browser/* group returns PLAIN JSON (no envelope).

## 1. Run an AI agent task — synchronous
POST /api/v1/ai-browser/execute            (Bearer px_live_)
Body (JSON; only "instruction" required):
  instruction   string   the goal, plain English ("task" accepted as an alias)
  startUrl      string   starting page ("url" accepted as an alias). The default engines
                         may not navigate there first: ALSO write it into the instruction
                         ("Go to https://... then ...")
  model         string   see model list below; OMIT for the default (gemini-3.5-flash)
  maxSteps      number   default 30
  maxCredits    number   hard AI-credit cap — always set this
  aiVision      string   screenshot-quality preset: adaptive (recommended), ultra-saver,
                         low, saver, efficient, balanced, detailed, enhanced, precision,
                         max-precision (cheaper presets = fewer image tokens)
  finalStepVision string one higher-quality look at the finished page (same slugs; "same" = off)
  outputSchema  object   JSON Schema for structured output (also accepted: output_schema)
  saveAsWorkflow bool    save the run as a replayable zero-AI-cost workflow (+ savedWorkflowName)
  recordSession bool     capture an MP4 of the run
  proxyEnabled  bool     residential proxy egress (+ proxyCountry, e.g. "us")
  agents        number   >1 = multi-agent swarm (see section 4)
Response data:
  { status: "completed" | "max_steps_reached" | "error",
    finalAnswer: string, steps: [...], creditsUsed: number,
    usage: { aiCredits, tokensIn, tokensOut, costUSD, turns },
    savedWorkflowId?: string, structuredOutput?: { data, error? } }
A sync request can run for minutes (server allows up to ~13 min). For anything beyond a
short task, use the async job mode below.

## 2. Async job mode — start / run / status / stop (for longer tasks)
Same endpoint, same auth. Four calls:
  POST {"mode":"start", ...same fields as sync...}   -> data: { jobId, sessionId, status:"running" }
  POST {"mode":"run","jobId":"..."}                  -> advances the job up to ~230s, returns status view
  GET  /api/v1/ai-browser/execute?jobId=...          -> cheap status view (does NOT advance the job)
  POST {"mode":"stop","jobId":"..."}                 -> data: { jobId, stopRequested:true, status }
Drive it by looping mode:"run" until the view says terminal:true (each call does real
work server-side). Status view fields:
  { jobId, sessionId, jobStatus: "running"|"awaiting_input"|"done"|"failed"|"stopped",
    status: "running"|"completed"|"error", terminal: bool, stepCount, steps,
    finalAnswer?, structuredOutput?, error?, errorCode?, savedWorkflowId?,
    creditsUsed, usage: { aiCredits, tokensIn, tokensOut, costUSD, turns } }

## 3. Models (the "model" field)
Default (omit the field): gemini-3.5-flash. An id outside this set returns 400
UNSUPPORTED_MODEL with the full supported list (a few legacy gemini-* aliases pass
validation but then fail at runtime — stick to this list). Verified list:
  Standard:      gemini-3.5-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-default
  Experimental:  gemini-3-flash-preview, gemini-3.5-flash-lite, model-a, model-a1,
                 model-b, model-j, model-k, model-k37
  Anthropic/OpenAI computer use: claude-sonnet-4-6, claude-opus-4-8, gpt-5.6-terra,
                 gpt-5.6-sol, gpt-5.5, gpt-6-astra, gpt-6.1-terra
  Generic vision: claude-fable-5-1-vision, claude-fable-5-vision, claude-opus-5-5-vision,
                 claude-opus-5-vision, claude-sonnet-5-5-vision, claude-sonnet-4-6-vision, gpt-6-astra-vision,
                 gpt-6.1-terra-vision, gpt-5.6-luna-vision, gpt-5.4-vision, gpt-4o-vision (Kimi models — kimi-k3-vision — see prompteryx.ai/legal/kimi for where their data is processed)
  Harness engines: modelc, model-d, model-d1, model-e, model-f, model-h, model-i
Notes: model-e / model-f answer 503 MODEL_NOT_CONFIGURED unless the deployment carries
their vendor keys; model-g needs deployment config (400 otherwise). For best
cost/quality just omit "model"; for the benchmark-champion harness use model-h (beta).

## 4. Multi-agent swarm
Add to the sync body:  agents: 2..N,  collaborate: true | false | "auto",
maxRunCredits: number (cap for the WHOLE swarm; maxCredits caps each lane).
collaborate true = agents share one work queue; false = N independent attempts;
"auto" = a planner picks. Response adds:
  swarm: { mode, status, agents: [per-lane results], coverage }
with merged top-level finalAnswer + usage summed across lanes.

## 5. Structured output
Pass outputSchema (JSON Schema). The finished run is projected into it and the response
carries structuredOutput: { data: <matching object> | null, error?: "why it could not" }.
structuredOutput is never silently absent when you supplied a schema.

## 6. Cloud-browser group (x-api-key: pcb_live_ — NOT the px_live_ key)
POST   /api/v1/cloud-browser/sessions   body: { viewport?, recordSession?, persistContext?,
       profileId?, proxy?, country?, sessionTimeoutMinutes?, captureDownloads? }
       -> { id, connectUrl, liveURL, status, startedAt }   (plain JSON, no envelope)
       Drive it yourself: chromium.connectOverCDP(connectUrl) from playwright-core.
GET    /api/v1/cloud-browser/sessions            -> { sessions: [...] } (recent, newest first)
GET    /api/v1/cloud-browser/sessions/:id        -> one session record
DELETE /api/v1/cloud-browser/sessions?sessionId=ID  (or DELETE /sessions/:id)
       -> { ok: true, proxyMB }   Close sessions when done — they meter minutes.
POST   /api/v1/cloud-browser/fetch   body: { url, format?: "text"|"markdown"|"html"|"links",
       waitUntil?, timeoutMs?, waitForSelector?, waitMs?, selectors?: [up to 10 CSS],
       screenshot?: bool, proxy?, country? }
       -> { url, finalUrl, title, status:"ok", content?|links?, extracted?, screenshotBase64? }
       One-shot: real Chromium renders the page (JS runs), you get token-efficient text.
POST   /api/v1/cloud-browser/search  body: { query, limit? (1-25), proxy?, country? }
       -> { query, results: [{ title, url, snippet }] }   (DuckDuckGo; no engine option)

## 7. Other verified platform endpoints (Bearer px_live_)
GET  /api/v1/auth/validate                     -> key + plan + credit balances (good first call)
GET  /api/v1/workflows?search=&limit=&offset=  -> { workflows, pagination }
POST /api/v1/workflows                         -> create { name, nodes:[{type,label,params}] } -> { workflowId }
POST /api/v1/workflows/:id/execute             -> body { variables?, options?, webhookUrl? }
                                                  -> 202 { execution: { id, status:"pending", statusUrl } }
GET  /api/v1/executions/:id                    -> { execution: { status, outputs, nodeResults } }
                                                  status: pending -> running -> COMPLETED|FAILED|STOPPED
                                                  (compare case-insensitively; runs start within ~1 min)
webhookUrl callback: POST { event:"workflow.completed"|"workflow.failed", data:{ executionId,
  status, outputs, ... } } signed X-Prompteryx-Signature-256: v1=HMAC-SHA256(whsec_ secret,
  "<X-Prompteryx-Timestamp>.<raw body>"); secret from the API Keys page; reject >5 min old.
GET  /api/v1/subscription                      -> plan, balances, usage
GET  /api/v1/schedules  (+POST, PATCH/DELETE /:id) -> server-side cron schedules (min interval 5 min)
GET  /api/v1/profiles   (+POST)                -> Chrome profiles

## Quick recipes (copy-paste; replace PX_KEY / PCB_KEY)
# 0) Verify the key works (free):
curl -s https://www.prompteryx.ai/api/v1/auth/validate -H "Authorization: Bearer PX_KEY"

# 1) One-shot agent task, cost-capped:
curl -s -X POST https://www.prompteryx.ai/api/v1/ai-browser/execute \
  -H "Authorization: Bearer PX_KEY" -H "Content-Type: application/json" \
  -d '{"instruction":"Open https://books.toscrape.com and report the title and exact price of the first book listed.","maxSteps":8,"maxCredits":10}'
# read .data.finalAnswer and .data.usage

# 2) Async job for a longer task:
curl -s -X POST https://www.prompteryx.ai/api/v1/ai-browser/execute \
  -H "Authorization: Bearer PX_KEY" -H "Content-Type: application/json" \
  -d '{"mode":"start","instruction":"Find the three cheapest science-fiction books on https://books.toscrape.com","maxSteps":25,"maxCredits":20}'
# then loop:  -d '{"mode":"run","jobId":"<jobId>"}'   until .data.terminal == true
# check:      curl -s "https://www.prompteryx.ai/api/v1/ai-browser/execute?jobId=<jobId>" -H "Authorization: Bearer PX_KEY"
# abort:      -d '{"mode":"stop","jobId":"<jobId>"}'

# 3) Structured output:
curl -s -X POST https://www.prompteryx.ai/api/v1/ai-browser/execute \
  -H "Authorization: Bearer PX_KEY" -H "Content-Type: application/json" \
  -d '{"instruction":"Open https://books.toscrape.com and read the first book listing.","maxSteps":8,"maxCredits":10,"outputSchema":{"type":"object","properties":{"title":{"type":"string"},"price":{"type":"string"}},"required":["title","price"]}}'
# read .data.structuredOutput.data

# 4) Swarm (three collaborating agents, whole-run cap):
curl -s -X POST https://www.prompteryx.ai/api/v1/ai-browser/execute \
  -H "Authorization: Bearer PX_KEY" -H "Content-Type: application/json" \
  -d '{"instruction":"Collect title and price of every book in the Travel category on https://books.toscrape.com","agents":3,"collaborate":true,"maxSteps":20,"maxCredits":10,"maxRunCredits":25}'

# 5) Cloud-browser session + one-shot fetch (pcb_live_ key!):
curl -s -X POST https://www.prompteryx.ai/api/v1/cloud-browser/sessions \
  -H "x-api-key: PCB_KEY" -H "Content-Type: application/json" -d '{}'
# -> { id, connectUrl, ... }  ... drive over CDP ... then:
curl -s -X DELETE "https://www.prompteryx.ai/api/v1/cloud-browser/sessions?sessionId=<id>" -H "x-api-key: PCB_KEY"
curl -s -X POST https://www.prompteryx.ai/api/v1/cloud-browser/fetch \
  -H "x-api-key: PCB_KEY" -H "Content-Type: application/json" \
  -d '{"url":"https://books.toscrape.com","format":"text"}'

## Not yet available — do not attempt (honesty notes, verified 2026-09-03)
- The npm package @prompteryx/sdk IS now published ("npm install @prompteryx/sdk") — a
  typed TypeScript wrapper over this same REST API. This raw HTTP guide still works fine.
- /api/v1/openapi.json covers only a subset of routes and mentions a "px_test_" key
  environment that DOES NOT EXIST (only px_live_ keys exist). Prefer this file.
- /api/v1/ai-browser/execute-stream is not SSE — do not stream; use async job mode.
- No API-key access to: API-key management (/api/v1/api-keys needs a signed-in browser
  session), Connect Hub integrations, workflow templates, custom nodes, usage
  time-series. Recordings: POST /api/v1/recordings only (no GET list).
- cloud-browser search has no "engine" option (always DuckDuckGo).

## Agent etiquette
- Cap every run: maxCredits <= 10 for tests. Prefer maxSteps <= 10 for smoke tests.
- Close cloud-browser sessions you create; stop async jobs you abandon (mode:"stop").
- Stay under 60 requests/min; poll async jobs no faster than every 5 seconds (GET), and
  drive them with mode:"run" calls rather than tight GET loops.