Agentdom

About

AgentDOM is the missing infrastructure layer between AI agents and the internet.

• Runtime (npm | pip): Routes any intent to the right execution layer — REST API, browser, desktop, or CLI
• Publisher SDK (npx agentdom-publisher): Expose your software to AI agents via .well-known/agentdom.json in minutes
• mcp-use bridge (npx agentdom-publisher mcpuse): Turn any manifest into a working Claude/Cursor MCP server instantly

Install

# No install needed — use npx directly
npx agentdom@latest onboard

# Or install globally
npm install -g agentdom

Python:

pip install agentdom
# or: pip install "agentdom[flask]"  pip install "agentdom[fastapi]"

Runtime: Node 18+

Quick start

# Step 1 — Guided setup wizard (one-time)
npx agentdom@latest onboard

# Step 2 — Connect integrations
npx agentdom@latest setup github.com      # Device flow — enter code at URL
npx agentdom@latest setup linear.app      # OAuth PKCE — browser opens once
npx agentdom@latest setup resend.com      # API key — paste once

# Step 3 — Run goals autonomously
npx agentdom@latest run "Create a Linear ticket for the iOS crash and assign to @alice"
npx agentdom@latest run "Send a weekly digest email to team@company.com via Resend"
npx agentdom@latest run "Open a GitHub PR for branch fix/login and request review from @bob"

# Health check
npx agentdom@latest doctor

How intent routing works

dispatch_intent("issues.create", args, "github.com")
        │
        ├─ 1. API manifest?    → instant REST call (< 200ms)
        ├─ 2. CLI tool?        → local command execution
        ├─ 3. Web UI only?     → Chrome CDP + DOM automation
        └─ 4. Desktop app?     → macOS Accessibility bridge

No screenshots. No coordinate clicking. No hand-written integrations.

Supported integrations (ready today)

Claude Desktop & Cursor (MCP)

claude mcp add agentdom -- node $(npm root -g)/agentdom/desktop-mcp-server.js

Or manually in claude_desktop_config.json:

{
  "mcpServers": {
    "agentdom": {
      "command": "node",
      "args": ["/path/to/agentdom/desktop-mcp-server.js"]
    }
  }
}

Embed in your own app

Express:

const { agentdom } = require('agentdom/express');

app.use(agentdom({
  host: 'api.myapp.com',
  capabilities: [{
    intent: 'todos.create',
    description: 'Create a todo item',
    args: { title: { type: 'string', required: true } },
    handler: async (req, body) => db.todos.create(body.title),
  }],
}));
// Serves: GET  /.well-known/agentdom.json
//         POST /api/agentdom/todos.create

Next.js (App Router):

// app/api/agentdom/[...intent]/route.js
import { AgentDOM } from 'agentdom/nextjs';

export const { GET, POST } = AgentDOM({
  host: 'myapp.com',
  capabilities: [{
    intent: 'users.invite',
    description: 'Invite a team member',
    args: { email: { type: 'string', required: true } },
    handler: async (req, body) => inviteUser(body.email),
  }],
});

Python (Flask / FastAPI):

from agentdom import AgentDOM

agent = AgentDOM(host="api.myapp.com")

@agent.capability("orders.ship", description="Ship a pending order",
                  args={"order_id": {"type": "string", "required": True}})
async def ship_order(args):
    return await shipping_service.ship(args["order_id"])

agent.register(app)   # Flask or FastAPI

Generate a mcp-use server from any manifest

# Generates a full TypeScript mcp-use server scaffold from any AgentDOM host
npx agentdom-publisher mcpuse --host=github.com --out=./github-mcp

# Python variant
npx agentdom-publisher mcpuse --host=github.com --lang=python

cd ./github-mcp && npm install && npm start
# → Inspector: http://localhost:3000/inspector
# → Connect to Claude Desktop or any MCP client

Publisher SDK — add AI agent support to your API

# Generate manifest from your OpenAPI spec
npx agentdom-publisher init --openapi=./openapi.json --host=api.myapp.com

# Validate locally
npx agentdom-publisher validate

# Verify live deployment
npx agentdom-publisher verify --host=api.myapp.com

# Test a real dispatch
npx agentdom-publisher test --host=api.myapp.com --intent=contacts.list --token=sk-...

# Submit to the AgentDOM public registry
npx agentdom-publisher submit --host=api.myapp.com

# Generate mcp-use server scaffold
npx agentdom-publisher mcpuse --host=api.myapp.com --out=./my-mcp-server

CI / Docker / serverless (headless)

# Export credentials as a portable base64 string (human runs this once)
agentdom wallet export --base64 --providers=linear.app,resend.com
# → AGENTDOM_WALLET_B64=eyJ3YWxsZXQi...

# Agent receives it at runtime — no human needed
docker run -e AGENTDOM_WALLET_B64=$AGENTDOM_WALLET_B64 \
           -e OPENROUTER_API_KEY=$OPENROUTER_API_KEY \
           your-agent-image

Policy engine

agentdom policy show                    # view current rules
agentdom policy set external=allow      # allow all external API calls
agentdom policy set send=prompt         # ask before emails/messages
agentdom approve <id>                   # approve a pending action
agentdom deny <id>                      # deny it

Deploy to AWS

./deploy-aws.sh sk-or-v1-YOUR_KEY

One command: ECS Fargate + ALB + CloudFront + S3 + Secrets Manager + CloudWatch. See AWS-DEPLOY.md.

All CLI commands

npx agentdom@latest onboard              Interactive setup wizard
npx agentdom@latest setup <provider>     One-time credential setup
npx agentdom@latest run "<goal>"         Execute a natural language goal
npx agentdom@latest doctor               Health check
npx agentdom@latest wallet list          Show stored credentials
npx agentdom@latest wallet export        Export for CI/Docker
npx agentdom@latest policy show          View policy rules
npx agentdom@latest approve <id>         Approve a pending action
npx agentdom@latest memory stats         Episodic memory stats
npx agentdom@latest launch               Browser REPL (interactive)
npx agentdom@latest --version            Print version
npx agentdom@latest --help               Show this help

Security

Credentials are stored in ~/.agentdom/wallet.json (mode 0600) or OS Keychain. Never logged, never sent to AgentDOM servers. See SECURITY.md.

License

MIT © Ragavendhra Machikatla