Agents talking to each other

An experiment comparing three ways browser-based agents can cooperate:

1. Direct WebMCP: the extension agent calls page tools itself.
2. Agent as a WebMCP tool: the extension delegates reasoning to the on-page agent.
3. A2A: the extension talks to the on-page or backend agent using A2A messages and task lifecycles.

Warning

This repository is for demonstration and educational purposes only. It does not represent a real product or production-ready service. The authentication, credential storage, persistence, security controls, agent behavior, and protocol coverage are deliberately simplified to make the examples easier to understand.

Do not use production API keys, personal information, confidential content, or other sensitive data with this demo, and do not expose it directly to the public internet. References to A2A compatibility describe only the tested subset documented below; they are not a security review, compliance claim, or certification.

Available under the MIT License.

Architecture

Component	Runs in	Role
On-page agent	Demo tab	Manages the notes board; exposed through WebMCP and experimental A2A-over-postMessage
Extension agent	Chrome side panel	Discovers page tools and chooses direct WebMCP, delegation, or A2A
Backend agent	Node server	A2A/HTTP agent with caller-scoped memos, meetings, contexts, and tasks

The page and extension use the same dependency-free multi-provider agent loop. The backend keeps provider keys server-side and cannot see the page DOM.

Supported models

The demo supports Anthropic, OpenAI, and Google Gemini. The site and extension each have a collapsed provider/model/key control. Keys are remembered per provider in browser storage; the backend reads its key from environment variables or .env.

Run the demo

Requirements:

• Node.js 20+
• Chromium with chrome://flags/#enable-webmcp-testing enabled
• A provider key for each agent you intend to run

Install dependencies:

npm install

Create .env, set one provider key, and optionally select the backend model:

ANTHROPIC_API_KEY=sk-ant-...
# OPENAI_API_KEY=sk-...
# GEMINI_API_KEY=AIza...
# AGENT_MODEL=claude-haiku-4-5

Start the server:

npm start

Then:

1. Open http://localhost:8787 and configure the on-page model.
2. In chrome://extensions, enable Developer mode and load extension/ unpacked.
3. Open the extension side panel and configure its model.
4. Discovery runs automatically. A retry button appears only when discovery fails.

After extension changes, click Reload in chrome://extensions.

Demo scenarios

Use the side panel channel selector or scenario buttons:

• Direct tools: summarize or edit notes through native WebMCP tools.
• Agent as tool: delegate a request to ask_page_agent through WebMCP.
• A2A to page: reach the same on-page agent with A2A v1-shaped JSON-RPC carried by the custom postMessage bridge.
• A2A to backend: reach the independent backend agent over HTTP/SSE.
• Input required: request a launch meeting without a day, then answer the backend's question against the same taskId.

The launch meeting is an invented demonstration scenario, not an official A2A example. Meetings and tasks exist only in backend memory.

Caller identities

The extension's Caller selector uses illustrative bearer tokens:

• Alice: demo-alice
• Bob: demo-bob

Backend contexts, tasks, memos, and meetings are bound to the authenticated caller. Alice cannot retrieve, resume, or cancel Bob's tasks. These tokens demonstrate the authorization boundary; they are not production credentials.

Standalone A2A client

The backend can be called without the website:

node examples/a2a-client.mjs "What memos do you have?"
node examples/a2a-client.mjs "Book a meeting about the launch" --stream
node examples/a2a-client.mjs "Friday" --task <taskId>

Use --token demo-bob to switch callers.

A2A compatibility

The backend implements a focused JSON-RPC/SSE subset of the released A2A v1.0.1 specification:

• Agent Card discovery through /.well-known/agent-card.json
• SendMessage
• SendStreamingMessage
• GetTask
• CancelTask
• input-required task resumption

Agent interfaces and the A2A-Version header use wire version 1.0, as required for the v1 protocol line. Automated interoperability tests use the official JavaScript v1 SDK, currently published as @a2a-js/sdk@1.0.0-beta.0.

This verifies the named subset; it is not certification of every A2A operation or transport. The browser-side postMessage pipe reuses A2A data and operation shapes but is not a standard A2A transport binding.

Verification

npm run build

The build checks generated extension copies and runs the model-free protocol and official-SDK interoperability tests.

When editing shared runtime files, synchronize their extension copies:

npm run sync:extension

Important limitations

• Browser keys are stored locally and sent directly to provider APIs.
• Backend state is in memory and disappears when the process stops.
• The postMessage capability hardens routing but cannot authenticate against another script already running in the same window.
• Built-in callers and bearer tokens are illustrative only.
• Production use would require real authentication, persistent storage, quotas, rate limits, broader protocol coverage, and operational hardening.

Troubleshooting

Problem	Fix
WebMCP unavailable	Enable chrome://flags/#enable-webmcp-testing and restart Chromium
No bridge in active tab	Focus the localhost:8787 tab and use Retry discovery
Missing provider key	Configure the selected provider in the relevant site, extension, or backend host
Extension behaves stale	Run npm run sync:extension, reload it in chrome://extensions, and reopen the panel
Port 8787 is occupied	Stop the existing process or configure a different server port consistently