Junyr now supports A2A: the open agent-to-agent protocol

Junyr implements the Agent2Agent (A2A) protocol — external AI agents can autonomously delegate business tasks to your Junyr agent, complementing the existing MCP connector. Discovered via an Agent Card, OAuth 2.1, scoped per company.

Junyr now supports A2A: the open agent-to-agent protocol

In short — Junyr now publishes a sovereign business agent that speaks the open A2A (Agent2Agent) protocol. Any external AI agent can discover it via a standard Agent Card, authenticate over OAuth 2.1, and delegate business questions to it — autonomously, with no human in the loop. It's the natural complement to the MCP connector: MCP connects your AI assistant (Claude, ChatGPT) to your tools; A2A lets other agents orchestrate Junyr as a peer.

---

🤖 What is A2A?

A2A (Agent2Agent) is an open agent-interoperability standard under the Linux Foundation. It defines how a client agent discovers a remote agent, sends it a task, and receives a result — over HTTP, using JSON-RPC 2.0, with optional SSE streaming.

The core building blocks:

• Agent Card — a JSON document published at /.well-known/agent-card.json describing the agent's identity, skills, endpoint, and authentication requirements.
• Task — a stateful unit of work with an id, a lifecycle (submitted → working → completed), and artifacts.
• Message / Part — a single turn of communication, made of content parts (text, file, data).
• Artifact — the concrete deliverable the agent produces.

🔌 A2A vs MCP — what's the difference?

Both protocols run simultaneously — you can use MCP and A2A together, with the same per-company isolation and the same confidentiality tiers.

🪪 Discovery + authentication

1. The client agent fetches Junyr's Agent Card: GET https://mcp.junyr.app/.well-known/agent-card.json.
2. The card declares the Business Q&A skill and the OAuth 2.1 (PKCE) security scheme — the same flow as the MCP connector.
3. The agent authenticates, obtains a token, then sends a JSON-RPC message/send request (or message/stream for streaming) to https://mcp.junyr.app/a2a.
4. Junyr replies with a Task and an Artifact carrying the answer, grounded in your company's connected graph (emails, contacts, projects, quotes, invoices, calendar).

🔒 Sovereignty and confidentiality

Junyr's A2A agent is scoped per company and gated by each company's confidentiality tier:

• Totale and Sécurisée → the agent rejects the task on the external A2A surface: no data leaves the perimeter to a third-party agent.
• Simple → the answer is produced over the company graph, always isolated.

Every input is filtered for prompt injection, and the surface is rate-limited. The Agent Card is static and generic — it never embeds any company-specific data.

❓ FAQ

Does A2A replace MCP?

No. They are complementary: MCP for human-driven use, A2A for autonomous agent-to-agent delegation. Both share the same authentication and isolation.

What data can the A2A agent query?

The connected business graph: emails, contacts, companies, projects, quotes, invoices, expenses, and calendar — read-only, in natural language (v1).

How is security handled?

OAuth 2.1 PKCE (the reused MCP flow), strict per-company isolation, a confidentiality gate (Totale/Sécurisée rejected on this surface), and anti-injection filtering.