Skip to main content
Coral Server integrates the MCP protocol to let AI agents collaborate. Each orchestrated agent runs with its own MCP server instance managed by Coral, exposing a standardized set of tools, resources, and communication channels over HTTP (SSE).

​
How Coral Server uses MCP

Every agent connected to a Coral Server communicates using the MCP protocol. When an agent is orchestrated, it is given a unique cryptographic credential (the agentSecret, provided via CORAL_AGENT_SECRET) and authenticates to its per‑agent MCP endpoint using this bearer token.

​
Standard Tools

Coral Server provides every agent with a default set of MCP tools for coordination (tool names are prefixed with coral_):
  • coral_create_thread — Create a new communication thread in the current session.
  • coral_send_message — Send a message to a specific thread (supports mentions).
  • coral_wait_for_message — Wait for the next message matching filters.
  • coral_wait_for_mention — Wait until the agent is mentioned.
  • coral_wait_for_agent — Wait for the next message from a specific agent.
  • coral_add_participant / coral_remove_participant — Manage thread participants.
  • coral_close_thread — Close a thread when a topic is complete.
  • coral_close_session — Optionally close the session (if enabled by plugins).

​
Resources

Coral Server exposes session‑related information through MCP resources:
  • Instructions (coral://instruction): Markdown guidance compiled from snippets describing how to use messaging, mentions, and waiting tools effectively.
  • State (coral://state): An agent‑visible snapshot of the current session state (e.g., active threads and recent messages).
For a detailed breakdown of what each resource contains and how they’re injected into agents, see Resources.

​
Integrating with your Application

If you’re building an application, you can integrate with Coral agents by:
  1. Providing Custom Tools: Define your own MCP tools and include them in the agentGraphRequest.customTools when creating a session. Coral Server will forward tool calls from agents to your specified HTTP endpoints.
  2. Puppeting Agents: Use the Puppet API to interact with agents as if you were another agent in the graph (e.g., create threads, send messages, manage participants).

​
Tool Definition Example

When defining a session, you can provide custom tools that agents can call: 
{
  "agentGraphRequest": {
    "customTools": {
      "query_database": {
        "transport": {
          "type": "http",
          "url": "https://your-api.com/query"
        },
        "schema": {
          "name": "query_database",
          "description": "Query the company's SQL database",
          "inputSchema": {
            "type": "object",
            "properties": {
              "query": { "type": "string" }
            },
            "required": ["query"]
          }
        }
      }
    }
  }
}
For more details on integrating with your application, see Integrating with your App.