AIR SDK

AIR SDK

Collective intelligence for agents. Stop your agents from guessing how to use the web. Let them know what to do. Think; hive mind. Let's make the open agent internet.

Install · Quick Start · Why · Agent Skill · Docs

---

Your browser agent wastes tokens and time every time it asks an LLM what to click. AIR SDK replaces guessing with knowing. This is an early research preview.

Up to 7,000x cost reduction and 280x faster vs frontier models. One function call. Zero code changes.

Benchmarks

Measured across 8 major domains, 40 cold API calls, cache disabled. LLM baseline priced at Frontier Model ($15/$75 per M tokens). AIR SDK at Scale tier ($149/mo, 250K executions).

The more complex the workflow, the more you save. LLM costs compound — every action adds to the conversation context, making each subsequent round trip more expensive. AIR's macro path is always one API call, regardless of how many steps the workflow has.

Methodology & raw data

LLM baseline: Agent inspects page DOM, reasons about which elements to interact with, generates tool calls. Frontier Model pricing ($15/M input, $75/M output).

• *1 action:* ~4,000 input + 800 output tokens x 2 round trips = ~9,600 tokens, ~$0.24, ~4s reasoning time.
• *10 actions:* Context grows with each action (conversation history accumulates). Average ~6,000 input + 1,000 output tokens x 2.5 round trips per action = 25 total LLM calls, ~175,000 tokens, ~$4.13, ~50s reasoning time.
• *Sonnet 4 comparison:* 5x cheaper per token — 1 action: ~$0.05, 10 actions: ~$0.83. Still 80–1,400x more expensive than AIR at Scale.

AIR SDK macro path: Pre-verified CSS selectors returned via API. Agent executes directly — no DOM inspection, no LLM reasoning. Cost = plan price / included executions.

Raw API latency (40 measurements, cache disabled):

• browse_capabilities: 180ms median, 148ms min, 421ms p95
• execute_capability: 178ms median, 146ms min, 333ms p95
• Combined: 354ms median, 304ms min, 1010ms p95

Cost per action by tier:

Capabilities discovered: 16 avg per domain (range: 3–39)

Last run: 2026-03-20 | Run benchmark yourself: npx tsx benchmark/run.ts

Prerequisites

• Node.js >= 18
• One of: Playwright, Puppeteer, or Browser Use (optional peer dependencies)

Install

npm install @arcede/air-sdk

npx @arcede/air-sdk init # saves key to ~/.config/air/credentials.json

Agent Skill (recommended for AI coding agents):

npx @arcede/air-sdk install-skill   # auto-configures Claude Desktop, Claude Code, Cursor, Windsurf, OpenClaw

Free tier included. No credit card. Make the network smarter, star the repo and share it ⭐

Lightweight Alternatives

Don't need the full SDK? We publish standalone packages for common use cases:

Quick Start

Wrap your existing Playwright page. Nothing else changes.

import { chromium } from 'playwright';

import { withAIR } from '@arcede/air-sdk/playwright';

const browser = await chromium.launch(); const page = await browser.newPage(); const smartPage = withAIR(page, { apiKey: process.env.AIR_API_KEY });

// Your existing code — now with site intelligence await smartPage.goto('https://example.com'); await smartPage.fill('#search', 'flights to tokyo'); await smartPage.click('.search-btn');

// Know what's possible on any site before writing automation const capabilities = await smartPage.air.listCapabilities('example.com');

Also works with Puppeteer and Browser Use:

// Puppeteer

import { withAIR } from '@arcede/air-sdk/puppeteer'; const page = withAIR(await browser.newPage(), { apiKey: process.env.AIR_API_KEY });

// Browser Use import { AIRPlugin } from '@arcede/air-sdk/browser-use'; const agent = Agent({ plugins: [new AIRPlugin({ apiKey: process.env.AIR_API_KEY })] });

Why

Browser automation is fragile. Selectors break. Sites change. Your agent has no idea what's possible until it's already on the page.

AIR SDK gives your agent three things:

How it works: Wrap your page with withAIR(). The SDK observes actions, resolves selectors, and preloads site capabilities. Your code doesn't change — it just works better.

Privacy-first: Input values, cookies, and PII are never sent. See our privacy docs →

Credential Management

API keys are stored at ~/.config/air/credentials.json with 0600 permissions (owner read/write only).

npx @arcede/air-sdk init       # Save your key (opens dashboard)

npx @arcede/air-sdk whoami # Show current key, source, and agent status npx @arcede/air-sdk logout # Remove key from all locations

Key resolution order: AIR_API_KEY env var → ~/.config/air/credentials.json → .env in current directory.

Agent Skill

Give your coding agent site intelligence with one command:

npx @arcede/air-sdk install-skill

This auto-detects Claude Desktop, Claude Code, Cursor, Windsurf, and OpenClaw, writes the MCP server config, and injects your API key. It installs @arcede/air-sdk globally for fast agent startup (~2s instead of ~60s with npx) and writes the absolute binary path to each config to avoid npx version caching issues. Restart your agent and it instantly has new tools.

Manual setup (Claude Desktop, Claude Code, Cursor, Windsurf, OpenClaw)

Requires global install: npm install -g @arcede/air-sdk. If not installed globally, replace "command": "air-sdk" with "command": "npx" and "args": ["--mcp"] with "args": ["-y", "@arcede/air-sdk", "--mcp"].

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{

"mcpServers": { "air-sdk": { "command": "air-sdk", "args": ["--mcp"], "env": { "AIR_API_KEY": "air_xxx" } } } }

Claude Code:

claude mcp add air-sdk -e AIR_API_KEY=your_key_here -- air-sdk --mcp

Cursor — add to ~/.cursor/mcp.json:

{

"mcpServers": { "air-sdk": { "command": "air-sdk", "args": ["--mcp"], "env": { "AIR_API_KEY": "air_xxx" } } } }

Windsurf — add to ~/.codeium/windsurf/mcp_config.json:

{

"mcpServers": { "air-sdk": { "command": "air-sdk", "args": ["--mcp"], "env": { "AIR_API_KEY": "air_xxx" } } } }

OpenClaw — add to ~/.openclaw/openclaw.json:

{

"mcpServers": { "air-sdk": { "command": "air-sdk", "args": ["--mcp"], "env": { "AIR_API_KEY": "air_xxx" } } } }

Tools available:

OpenAI Skill

AIR SDK is also available as an OpenAI hosted shell skill for gpt-5.4 and gpt-5.4-mini:

# Package and upload

cd air-sdk/openai-skill zip -r /tmp/air-sdk-skill.zip . curl -X POST 'https://api.openai.com/v1/skills' \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -F 'files=@/tmp/air-sdk-skill.zip'

Then use in the Responses API:

{

"model": "gpt-5.4-mini", "tools": [{ "type": "shell", "environment": { "type": "container_auto", "skills": [{ "type": "skill_reference", "skill_id": " " }], "network_policy": { "type": "allowlist", "allowed_domains": ["agentinternetruntime.com", "api.agentinternetruntime.com"], "domain_secrets": [ { "domain": "api.agentinternetruntime.com", "name": "AIR_API_KEY", "value": " " }, { "domain": "api.agentinternetruntime.com", "name": "Authorization", "value": "Bearer " } ] } } }] }

Requires api.agentinternetruntime.com in your org network allowlist.

Configuration

withAIR(page, {

apiKey: process.env.AIR_API_KEY, // Required cacheEnabled: true, // Local capability cache (default: true) cacheTTLMs: 1_800_000, // Cache TTL in ms (default: 30 min) debug: false, // Log debug info (default: false) });

Machine Payments