BYO A2A Agents & MCP Servers
CAIPE is the orchestration layer — plug in your own A2A-compatible agents and MCP servers via the supervisor agent registry, dynamic-agents seed config, or Docker Compose. Your protocols, your tools.
Quick links: Supervisor Chart · Dynamic Agents Chart
A2A Agent Registry — Supervisor Helm Chart
Register external A2A agents with the supervisor via multiAgentConfig.
| Key | Purpose |
|---|---|
multiAgentConfig.agents | Explicit list of external A2A agent endpoints to register |
multiAgentConfig.protocol | a2a (peer-to-peer) or slim (hub-based) transport |
multiAgentConfig.port | Port advertised to peer agents for inbound connections |
Auto-discovery mode: leave agents empty and the supervisor discovers peers on the network. Works with any A2A-compatible agent — CAIPE agents, custom FastAPI services, or third-party implementations.
MCP Servers — Dynamic Agents Seed Config
Register external MCP server endpoints at chart install time using seedConfig.mcp_servers.
- Each entry specifies name, URL, and optional auth headers
- MCP servers are available to all custom agents in the dynamic-agents service
- Hot-register new MCP servers from the Skills Gateway UI without redeployment
- Supports HTTP/SSE MCP transport — compatible with FastMCP and any spec-compliant server
Docker Compose — A2A Transport
| Env Var | Purpose |
|---|---|
A2A_TRANSPORT | p2p for peer-to-peer or slim for hub-based routing |
NEXT_PUBLIC_A2A_BASE_URL | Point the UI at your custom supervisor or gateway endpoint |
Add any external A2A agent as a new service in docker-compose.yaml with the shared network. Each agent service exposes its A2A endpoint; the supervisor discovers and routes to it automatically.
Supported Protocols
| Protocol | Description |
|---|---|
| A2A (Agent-to-Agent) | Google-led open protocol for agent interoperability |
| MCP (Model Context Protocol) | Anthropic-led standard for tool and resource exposure |
| AG-UI / SSE | Real-time event streaming across all agent types |
| SLIM | Hub-based routing for firewall-friendly deployments |
| OpenAI-compatible | Any agent exposing /v1/chat/completions works |
Bring Your Own LLM
Any OpenAI-compatible endpoint works — Claude (via Anthropic), GPT-4o, Gemini, Llama, Mistral.
- Configure per-agent model endpoints in
llmSecretsorseedConfig.models - LiteLLM proxy support — route to any provider through a single unified endpoint
- Switch models without redeploying — update via Helm values or the UI
Backstage & External Integrations
- Agent Forge Backstage plugin — surface CAIPE agents directly in your Internal Developer Portal
- Slack Bot / Webex Bot — expose the full supervisor as a conversational interface
- CLI access — invoke any agent from the terminal via the CAIPE CLI
- REST API — integrate agent execution into CI/CD pipelines or custom dashboards