Open protocol · SDK · Station · CLI

Give your agents apps,
not just tools.

Agent Apps are applications designed for AI agents. They give agents tools to act, interfaces to present information, and a full runtime to manage their lifecycle: install, configure, authenticate, run, and scale.

npm create @agentapps

Or use the +80 that exist↓

+80 apps ready to installOpen protocol · MITBuilt by Pablo López→

Works today in

Teros

Claude Code

Claude Desktop

Cursor

Windsurf

VS Code

The gap

A tool call returns text.
An Agent App returns a product.

Agents have had function calling (the model picking one of your functions and filling in its arguments) for years. What they never had was somewhere to put a screen, keep state, hold a credential, or hand work to another app. That is the whole point of this protocol.

With tools alone

·Text in, text out. Every result is a paragraph the agent has to narrate.
·No interface. Tables, charts and forms get flattened into prose.
·No memory between calls. The agent re-establishes context every time.
·No lifecycle. Nothing to install, configure or authenticate against.
·No way for one integration to talk to another.
·One process for everyone, or nothing at all.

With Agent Apps

01 Typed tools with parameters, permissions and structured results.
02 Real interfaces, rendered natively by the host: cards, tables, charts, forms, images.
03 Platform-provided storage: the app stores data, the platform owns it.
04 A managed lifecycle with OAuth or API-key auth and health checks.
05 Intents, so one app can invoke another's capabilities.
06 standalone, cluster or on-demand. One instance per user if you need it.

Gateway

You don't have to adopt
anything to start.

Agent Apps speak their own protocol. Gateway is the bridge that re-exposes them as MCP tools. MCP, the Model Context Protocol, is the standard your client already speaks to discover and call tools, so it sees all +80 apps as native tools. No fork, no plugin, no change to a single app.

Your apps

+80 Agent Apps

Notion, GitHub, Playwright, Calendar. Each one a manifest, tools and permissions.

protocol v2.0

Serving them

Station

One endpoint. Spawns, proxies, restarts and idles instances.

translates

The bridge

Gateway MCP

Speaks the protocol on one side, MCP on the other. That is the whole trick.

mcp / stdio

Where you work

Your client

Teros, Claude Code, Cursor, Windsurf, VS Code, Zed. Unmodified.

Two commandszsh

# 1. start the station with every installed app$ agentapps serve 4100 station listening · 79 apps · 10 running# 2. point your MCP client at it$ claude mcp add agent-apps \ -e STATION_URL=http://localhost:4100 \ -- node /path/to/gateway-mcp/dist/index.js ✓ agent-apps connected# every tool of every installed app is now callable from the client$

Works today in

Teros

Claude Code

Claude Desktop

Cursor

Windsurf

VS Code

Gateway MCP setup→

What it looks like on the wire

Ask a question.
Watch what the app does.

You say "am I free Thursday afternoon?". The agent picks the tool. The app answers with structured data, not prose. Later, someone else moves a meeting and the app tells the agent, unprompted.

app.google.calendarsession plane

# the agent chose the tool and the arguments. you never typed this.→ tool_call get_freebusy { "calendars": ["primary"], "timeMin": "2026-08-13T12:00Z", "timeMax": "2026-08-13T18:00Z" }← tool_result ok · 218 ms { "busy": [ { "start": "…T15:00Z", "end": "…T16:30Z" } ] }# a structured result, so the host can render a calendar strip instead of a paragraph# twenty minutes later, a colleague moves a meeting in Google Calendar← subscription_event calendar_updated { "action": "updated", "eventId": "3f9a…", "title": "Design review" }$

What a plain tool would have returned

01A paragraph the agent has to parse and re-narrate.
02Nothing when the calendar changes. You ask again, or you don't find out.
03A token per call, re-authenticated, with no session to hold.

What the Agent App returned

01Typed JSON the host can render as a real interface.
02A calendar_updated event, pushed, without being asked.
03An OAuth session the platform owns and refreshes for it.

The catalogue

+80 apps
already written.

The protocol is not a proposal waiting for adopters. The repo ships +80 apps across 22 categories, every one of them declaring its tools and its permissions in a manifest that validates today.

Browse all +80→

The catalogue · the lit ones are migrated and running

By category

What an Agent App provides

Six things you'll use.
Six more when you need them.

An Agent App can be a single tool, or a full product with screens, storage, app-to-app calls and its own embedded AI agent. Everything below is declared in the manifest and enforced by the protocol. None of it is inferred.

Tools

Functions an AI agent can call, with typed parameters, permissions and structured results.

Interfaces

Visual UI rendered natively by the host: cards, tables, charts, forms, images. Falls back to text or markdown.

Lifecycle

Install, configure, authenticate (OAuth / API key), start, stop, health checks.

Events

Real-time notifications from the app to the agent: webhooks, state changes, alerts.

Permissions

Granular grants over which tools, which data and which UI surfaces. Requested by the app, approved by the user. Tool hints like destructiveHint are declared, never inferred.

Intents

App-to-app communication. One app can invoke another's capabilities, through a binding the user controls.

And when the app grows into a product

ScreensStorageApp AgentsConversation hooksVolumes & FilesTelemetry

Architecture

Transport-agnostic
by design.

A host connects to a single station endpoint. The station spawns, routes, restarts and idles your app instances. Your code never deals with transports.

Host

The AI platform

An IDE, assistant or agent runtime that discovers, installs and uses Agent Apps on behalf of its users.

Station

Lifecycle & routing

One endpoint for hosts. Spawns, proxies, restarts and idles instances. Four runtime adapters, crash detection, graceful shutdown and sharding.

App instance(s)

standalone · 1 sharedcluster · N replicason-demand · 1 per user4 runtime adapters

01InstallonInstall · grants consent

02Configuresystem + user secrets

03AuthenticateOAuth2 · PKCE · API key

04Runspawn · proxy · idle

05RefreshonRefreshToken

06UpgradeonUpgrade · onUninstall

Agent App

Your code. A self-contained application with tools, interfaces and logic. Runs standalone, cluster or on-demand.

Station

Manages the app's lifecycle: spawns, routes, restarts and idles instances. Exposes a single endpoint for hosts to connect to. Transport-agnostic.

Host

The AI platform (IDE, assistant, agent runtime) that discovers, installs and uses Agent Apps on behalf of its users.

Gateway

A bridge that exposes Agent Apps to tool protocols like MCP, so they work in Teros, Claude Code, Cursor, Windsurf, VS Code and Zed with no changes.

Building one yourself

Scaffold, run, test.
All of it local.

If you want to write one instead of using one: a REPL and CLI to run apps locally, a test harness for your suite, and a scaffolder that gets a new app to its first tool call in a single command.

zsh

# Requires Node >= 20, pnpm and git$ git clone git@github.com:agentapps/agentapps.git$ cd agentapps$ ./install.sh installing dependencies… building @agentapps/cli… symlinked `agentapps` → /usr/local/bin$

The manifest

One file says
what the app is.

Every app ships an agentapp.json that has to pass safeParseAppManifest() from the protocol package. If it does not validate, it does not install. No optional-field drift, no undocumented behaviour.

Required fields

idStable identifier, e.g. app.notion
versionSemver, independent of the protocol
nameDisplay name
descriptionWhat the app does
tagline120 characters, hard limit
authorName and email
logoSVG, or PNG at 1024×1024 minimum
logoMonoSingle-tint SVG the host recolours
termsURL to the terms that apply
categoryOne of 22 in use today
entrypointModule path and export
runtimeTransport, command, arguments

apps/notion/agentapp.jsonvalid

"id": "app.notion",
"version": "2.0.0",
"tagline": "Manage Notion pages, databases, and content",
"license": "MIT",
"category": "productivity",

"entrypoint": {
  "type": "module",
  "path": "src/app.ts",
  "export": "factory"
},
"runtime": {
  "transport": "stdio",
  "command": "tsx"
},
"auth": {
  "type": "oauth2",
  "scopes": ["read_content", "update_content"],
  "pkce": false,
  "systemSecrets": ["CLIENT_ID", "CLIENT_SECRET"],
  "userSecrets": ["ACCESS_TOKEN", "REFRESH_TOKEN"]
}

Packages

Eight of them.
Take only what you need.

Take the protocol alone to implement your own host. Take the SDK to write an app. Take the station to run them.

protocolZod schemas, both planes, error codes sdkAuthor an app: tools, hooks, events, UI stationLifecycle, proxy, runtime adapters hostWhat a platform embeds to consume apps execution-hostStandalone deployable for instances cliREPL and one-shot, run apps locally testingHarness plus conformance suite createScaffold, dev mode, validate, pack

Elsewhere

Two more
ways in.

Gateway · setup

Connect the client you already use

Serve your installed apps, then point Teros, Claude Code, Cursor or Windsurf at them. Six clients, one config block each.

Read it→Creator

Pablo López

Founder, startup advisor, and the person building this. Software architecture, cybersecurity and scaling systems.

Read it→

Get started

Your first Agent App is
one command away.

Scaffold it, run it in the REPL, then expose it to Teros, Claude Code, Cursor, Windsurf, VS Code or Zed through Gateway MCP, without changing a line.

npm create @agentapps

Read the docs↗Star on GitHub↗

Give your agents apps,not just tools.

A tool call returns text.An Agent App returns a product.

With tools alone

With Agent Apps

You don't have to adoptanything to start.

Ask a question.Watch what the app does.

What a plain tool would have returned

What the Agent App returned

+80 appsalready written.

Six things you'll use.Six more when you need them.

Tools

Interfaces

Lifecycle

Events

Permissions

Intents

Transport-agnosticby design.

Agent App

Station

Host

Gateway

Scaffold, run, test.All of it local.

One file sayswhat the app is.

Eight of them.Take only what you need.

Two moreways in.

Connect the client you already use

Pablo López

Your first Agent App isone command away.

Give your agents apps,
not just tools.

A tool call returns text.
An Agent App returns a product.

You don't have to adopt
anything to start.

Ask a question.
Watch what the app does.

+80 apps
already written.

Six things you'll use.
Six more when you need them.

Transport-agnostic
by design.

Scaffold, run, test.
All of it local.

One file says
what the app is.

Eight of them.
Take only what you need.

Two more
ways in.

Your first Agent App is
one command away.