local-first · zero config · no cloud

AirPlay for AI agents

Your agents discover each other on your network. Zero config. No cloud. Just run npx agentdex.

Get in Touch See How It Works
What Actually Happens
1

You spin up multiple agents on your network

Claude Code on your laptop, an Ollama model on your desktop, a small agent on a Raspberry Pi. All on the same WiFi or nearby over Bluetooth.

Claude Code Ollama · llama3 Pi agent
2

Run npx agentdex and agents broadcast automatically

Each agent announces its name and skills over WiFi (mDNS) and Bluetooth (BLE mesh). No registry, no account, no server.

Broadcasting on WiFi · mDNS BT · BLE mesh
3

Your primary agent discovers specialists by skill

Need a code review? AgentDex finds the best match on your network and delegates automatically via the A2A protocol.

const agents = await dex.discover() // 3 agents found on network const result = await dex.delegate('review this PR') // routed to Code Reviewer (WiFi, 12ms)

Results flow back. Zero configuration.

Your agents collaborate. Results stream back to your primary agent. You never configured a URL, a key, or a registry entry.

✓  Task complete. No config required
How It Works

One command. Instant mesh.

AgentDex turns any function into a discoverable agent. Run, discover, and delegate. All on your own hardware.

1

Run

One command and your agent starts broadcasting its skills over both WiFi and Bluetooth simultaneously.

$ agentdex \ --name "My Agent" \ --skills "code-review"
2

Discover

Agents appear like AirPlay devices, automatically visible on your network without any setup or central directory.

# WiFi: mDNS service discovery # BT: BLE mesh advertisement # No server. No account.
3

Delegate

Skill-based routing via the A2A protocol. Pass files, tasks, or any structured data between agents.

# Files, tasks, anything dex.find('code-review') .send({ file: './pr.diff' })
Provider API

Turn any function into a discoverable agent

Three lines to register a skill. AgentDex handles all the broadcasting, discovery, and protocol negotiation for you.

TypeScript agent.ts
import { agent } from 'agentdex'

agent('code-reviewer', {
  skills: ['code-review', 'security-audit'],
  handle: async (task) => {
    return await myModel.review(task.input)
  }
})
Three lines to turn any function into a discoverable agent.
Two Transports

WiFi and Bluetooth, out of the box

AgentDex discovers agents over both transports simultaneously. Whether you're on a fast local network or working air-gapped, your agents can find each other.

📶

WiFi

mDNS · Bonjour

Covers your entire local network. High bandwidth HTTP delegation, perfect for large files, long-running tasks, and fast round-trips.

  • Covers your entire LAN automatically
  • High bandwidth: files, diffs, large payloads
  • HTTP delegation via A2A protocol
  • Latency under 5ms on a local network
🔌

Bluetooth

BLE mesh

Proximity and air-gapped scenarios. Works offline and behind firewalls using multi-hop BLE mesh routing.

  • Works fully offline, no WiFi needed
  • Multi-hop BLE mesh networking
  • Air-gapped environments and secure networks
  • Low power, runs on embedded devices
Dashboard

See every agent on your network at a glance

The macOS app shows you all discovered agents, their transports, latency, and available skills. All in real time.

AgentDex | Local Agent Network
CC
Code Reviewer
MacBook Pro · 192.168.1.42
WiFi 12ms
code-review security-audit refactor explain
247
Tasks
12ms
Latency
99%
Uptime
CLI

Find and delegate from the terminal

The AgentDex CLI lets you search agents by skill, send tasks, and pipe output, all from your existing workflow. No GUI required.

Terminal
$ agentdex find "code review"
   Code Reviewer     WiFi    12ms
   Ollama · llama3   WiFi    28ms

$ agentdex send "review this PR" --file ./diff.md
  Routing to: Code Reviewer (WiFi, 12ms)
  Task sent ✓

$ agentdex list
   Code Reviewer   code-review, security-audit
   Ollama llama3   summarize, translate
   Pi Agent        sensor-read, gpio-control
Potential Use Cases

Where AgentDex fits in

The agent ecosystem is growing fast. AgentDex sits at the discovery layer, making agents findable so protocols like A2A can do the rest.

🌐

LAD–A2A local catalogue

LAD–A2A is the emerging protocol for discovering A2A agents on local networks. AgentDex can serve as the smart local catalogue node, not just answering "who's here?" but adding semantic search, skill matching, and quality signals on top of raw discovery.

LAD–A2A semantic search catalogue
🔗

A2A mesh node

The A2A community is proposing small, fast catalogues inside every local mesh that gossip changes into an eventually consistent global directory. AgentDex is built to be exactly that node, aggregating _a2a._tcp.local advertisements and providing ranked, searchable results.

A2A mesh gossip protocol agent catalogue
🔭

OpenAgents federation

OpenAgents builds persistent agent networks across the internet using MCP and A2A. AgentDex handles the local LAN layer with fast, zero config discovery on your network, while federating with OpenAgents for cross network reach when agents need to go beyond the LAN.

OpenAgents MCP federation
🤖

Multi model agent pipelines

Claude writes the code, Ollama reviews it for security, a local linter checks style. Each agent discovers the others and delegates automatically. No orchestrator, no hardcoded URLs, no central registry.

Claude Ollama auto routing

Recursive task DAGs

Complex workflows decompose into subtasks that fan out across your network. AgentDex routes each DAG node to the best local agent by skill match. A task that needs "code review" automatically finds the nearest capable agent.

task DAGs skill routing fan out
🔒

Air gapped agent mesh

Compliance sensitive or classified environments where nothing touches the internet. AgentDex's BLE transport creates a proximity mesh between nearby devices. Agents collaborate over Bluetooth without WiFi, without a router, without any cloud dependency.

BLE mesh offline zero infrastructure
Ecosystem

The missing piece in agent infrastructure

Three protocols, three jobs. AgentDex fills the discovery gap.

AgentDex

Discovery

"Who's here?"

A2A

Communication

"What can you do?"

MCP

Integration

"How do I use it?"

🏢 Office LAN
"Book conference room 4B for 2pm"
Routed to: facilities-agent (WiFi, 8ms)
via Facilities Agent
🏠 Home Lab
"Transcribe this meeting recording"
Routed to: whisper-agent (WiFi, 3ms)
via Mac Mini / Whisper
💻 Dev Network
"Review this PR for security issues"
Routed to: code-reviewer (WiFi, 12ms)
via Ollama / CodeLlama
Under the Hood

Built on open standards

AgentDex combines proven protocols: mDNS for discovery, A2A for delegation, JSON-RPC for transport. No proprietary lock-in.

mDNS Discovery Transport

Agents register as _a2a._tcp.local services via multicast DNS, the same protocol that powers AirPlay, Chromecast, and printer discovery.

service: _a2a._tcp.local name: code-reviewer port: 3001 txt: skills=code-review,security

A2A Protocol Delegation

Google's Agent-to-Agent protocol. Each agent exposes a POST /a2a endpoint accepting JSON-RPC 2.0 method calls for task delegation.

{ "jsonrpc": "2.0", "method": "tasks/send", "params": { "task": { "input": "review this" } } }

Agent Cards Identity

Each agent serves a /.well-known/agent.json card describing its capabilities, skills, and supported protocols.

{ "name": "code-reviewer", "skills": ["code-review"], "url": "http://local:3001", "a2a": "/a2a" }
Component Technology Purpose
Discovery mDNS / Bonjour (RFC 6762) Zero-config service discovery on local networks
BLE Transport Bluetooth Low Energy mesh Proximity discovery and air-gapped environments
Delegation A2A over JSON-RPC 2.0 Structured task delegation between agents
Identity Agent Cards (.well-known) Capability advertisement and protocol negotiation
Server Hono (lightweight HTTP) A2A endpoint and dashboard serving
Registry In-memory, merge-based Live agent state from multiple transports
Router Skill matching + outbox Best-match delegation with retry and expiry

Under active development

AgentDex is being built in the open. The repo will be shared soon. If you're interested in local-first agent discovery, reach out.

Get in Touch Read the Docs

🔒 Coming Soon Source will be shared publicly once the core is stable.