Agent Actions Protocol (AAP)

Open Protocol for Structured UI-to-Agent Interactions - When AI agents render HTML interfaces, users interact with them. AAP captures those interactions as structured JSON records, stores them immutably, and notifies the agent.

The Problem

AI agents increasingly render rich HTML interfaces — dashboards, forms, reports, interactive tools. Users interact with these UIs, but the interactions have no standard way to flow back to the agent.

Current Approach	Problem
console.log scraping	Fragile, no structure, lost on refresh
Polling the DOM	Complex, race conditions, not sandboxed
Custom WebSocket per page	Over-engineered, security nightmare

The Solution

AAP defines a minimal protocol: one bridge script, one message format, one storage pattern.

Agent writes HTML  -->  Dashboard renders in iframe  -->  User clicks button
                                                              |
Agent reads JSON   <--  File written + notification   <--  postMessage to parent
                                                              |
                                                         HTTP POST to API

Quick Example

In the agent's canvas HTML:

<button onclick="maestro.send('click', 'approve-btn', { approved: true })">
  Approve
</button>

What the agent receives (JSON file):

{
  "id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "timestamp": "2026-05-18T15:30:00.000Z",
  "canvasFile": "reports/dashboard.html",
  "action": "click",
  "element": "approve-btn",
  "data": { "approved": true },
  "summary": "User click 'approve-btn' on reports/dashboard.html with data: {approved: true}"
}

Standard Action Vocabulary

Action	Use case
click	Button press, link activation
submit	Form submission
change	Input value changed
select	Option selected from dropdown/list
toggle	Boolean switch toggled
dismiss	Modal/notification dismissed
navigate	In-canvas navigation
custom	Application-specific action

Custom actions beyond this vocabulary are allowed — the action field is freeform.

Specification

The full protocol specification is in the spec/ directory:

• 01-introduction.md — Design principles and overview
• 02-bridge.md — Bridge script injection and API
• 03-messages.md — Action message format
• 04-storage.md — Interaction record format and file storage
• 05-transport.md — Transport and API endpoints
• 06-security.md — Security considerations
• 07-implementers.md — Implementer checklist

Or read the complete spec in one file: PROTOCOL.md

Relationship to AMP & AID

AAP is independent — it does not require AMP or AID. All three are complementary:

Protocol	Direction	Purpose
AAP	User → Agent	Structured UI interactions
AMP	Agent → Agent	Signed messages and routing
AID	Agent identity	Cryptographic authentication

All three share the same agent directory structure (~/.aimaestro/agents/<id>/).

Claude Code Plugin

This repo includes a canvas-actions skill for Claude Code agents. When installed, agents automatically know how to create canvas pages, embed interactive data, and process user interactions.

Installation

The skill is distributed via the AI Maestro plugin:

# Install AI Maestro plugin (includes AAP + AMP + AID skills)
git clone https://github.com/23blocks-OS/ai-maestro-plugins.git
cd ai-maestro-plugins
./build-plugin.sh --clean
./install-plugin.sh -y

Or add AAP as a source in your own plugin manifest:

{
  "name": "agent-actions",
  "type": "git",
  "repo": "https://github.com/agentmessaging/agent-actions.git",
  "ref": "main",
  "map": {
    "skills/canvas-actions": "skills/canvas-actions"
  }
}

What the Skill Teaches Agents

• When to create a canvas (proactive triggers like "show me", "visualize", "dashboard for")
• How to write data-driven interactive HTML with embedded JSON
• Where to store canvas files (~/.aimaestro/agents/<id>/canvas/)
• How to process [CANVAS] notifications and read interaction records
• Best practices for the maestro.send() bridge API

Reference Implementation

AI Maestro is the reference implementation of AAP. It includes the bridge script injection, canvas rendering, interaction storage, and agent notification.

Roadmap

Version	Scope
v1.0 (current)	One-way canvas interactions (user → agent)
v1.1	Bidirectional — agents push updates back to canvas
v1.2	Interaction acknowledgment — agent confirms receipt
v2.0	Agent-defined UI components (widgets, forms, controls)

License

This specification is released under the MIT License.

Copyright 2026 Juan Pelaez / 23blocks.