<p align="center"> <h1 align="center">AIR SDK</h1> <p align="center"> 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. </p> </p>

<p align="center"> <a href="#install">Install</a> · <a href="#quick-start">Quick Start</a> · <a href="#why">Why</a> · <a href="#agent-skill">Agent Skill</a> · <a href="https://agentinternetruntime.com/docs/sdk">Docs</a> </p>

---

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).

| Scenario | LLM DOM Reasoning (Frontier Model) | AIR SDK (Macro Path) | Savings | |----------|-------------------------------|---------------------|---------| | 1 browser action | ~$0.24, ~4s, ~10K tokens (2 LLM round trips to inspect DOM and act) | $0.0006, 178ms, 0 tokens | 400x cost reduction, 22x faster | | 10 browser actions | Up to ~$4, ~50s, ~175K tokens (25 LLM round trips as context grows) | $0.0006, 178ms, 0 tokens | Up to 7,000x cost reduction, 280x faster |

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.

<details> <summary>Methodology & raw data</summary>

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: | Tier | Price | Included | Effective cost | |------|-------|----------|---------------| | Free | $0 | 1,000/mo | $0 | | Pro | $49/mo | 25,000/mo | $0.0020 | | Scale | $149/mo | 250,000/mo | $0.0006 |

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

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

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:

| Package | Install | Use case | |---------|---------|----------| | @arcede/air-mcp | npx @arcede/air-mcp | MCP server only — add AIR tools to Claude Code, Cursor, or Windsurf without the full SDK | | @arcede/air-cli | npx @arcede/air-cli | Terminal CLI — extract data and query capabilities from the command line | | @arcede/extract | npm i @arcede/extract | Typed TypeScript client for the Extract API only | | @arcede/capabilities | npm i @arcede/capabilities | Typed TypeScript client for the Capability API only |

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:

| What | How | |------|-----| | Site intelligence | Know what actions are possible on any website — search, purchase, login, browse — before your agent navigates there. | | Resilient selectors | When a selector fails, the SDK automatically resolves fallback alternatives so your automation doesn't break. | | Continuous learning | The platform gets smarter over time, improving reliability and coverage across the web. |

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.

<details> <summary>Manual setup (Claude Desktop, Claude Code, Cursor, Windsurf, OpenClaw)</summary>

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" }
    }
  }
}

</details>

Tools available:

| Tool | Description | |------|-------------| | extract_url | Extract structured data from any URL — JSON-LD, RSS/Atom feeds, JSON APIs, SPAs. Meta-only results cost 0 credits. | | browse_capabilities | Discover what actions can be automated on a website. Returns confidence scores, tiers, selectors, and universal patterns. | | execute_capability | Get a structured execution plan with CSS selectors, fallbacks, and pattern-matched guidance. | | report_outcome | Report execution results with optional browserObservations to improve collective intelligence. | | extract_content | Extract text and structured sections from local files or remote file URLs such as PDFs, DOCX, PPTX, XLSX, CSV, TXT, MD, HTML, and EML. |

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": "<your_skill_id>" }],
      "network_policy": {
        "type": "allowlist",
        "allowed_domains": ["agentinternetruntime.com", "api.agentinternetruntime.com"],
        "domain_secrets": [
          { "domain": "api.agentinternetruntime.com", "name": "AIR_API_KEY", "value": "<key>" },
          { "domain": "api.agentinternetruntime.com", "name": "Authorization", "value": "Bearer <key>" }
        ]
      }
    }
  }]
}

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 Protocol

AIR SDK supports x402 micropayments, enabling your agent to pay for premium site access without pre-registration:

const smartPage = withAIR(page, {
  apiKey: process.env.AIR_API_KEY,
  mpp: {
    enabled: true,
    paymentMethod: 'stripe_spt',
    credential: 'spt_...',
    maxPerRequest: 0.50,
    maxPerSession: 5.00,
    autoApprove: true,
  },
});

Examples

See examples/agents/ for working agent examples:

Research Agent — Multi-source research with capability-aware navigation
Price Comparison — Cross-site price monitoring with resilient selectors
Capability Scout — Discover and map capabilities across domains
Site Monitor — Track site changes with resilient selectors

Contributing

See CONTRIBUTING.md. PRs welcome.

License

MIT

AIR SDK

Benchmarks

Prerequisites

Install

Lightweight Alternatives

Quick Start

Why

Credential Management

Agent Skill

OpenAI Skill

Configuration

Machine Payments Protocol

Examples

Contributing

Links

License

Score

Proud of your score? Add this badge to your README.

AIR SDK FAQ

How do I install AIR SDK?

Is AIR SDK an official MCP server?

Related MCP servers

Related guides