MCP Integration

import { Aside } from ‘@astrojs/starlight/components’;

Overview

Cyntex embeds an MCP server in-process (HTTP transport). An AI agent connects via the standard MCP protocol and gains real-time operational access to the running Cyntex instance.

AI Agent (Claude / GPT-4o …)
    │  MCP protocol (HTTP)
    ▼
Cyntex MCP server (in-process)
    │  direct call
    ▼
Control Core (CRUD + lifecycle + read-only runtime)

Key design: the MCP server contains no LLM — the model is on the user’s agent side; Cyntex only provides the operational interface.

Connecting

Server URL: http://localhost:7778/mcp
Auth: Bearer <token>

A localhost bootstrap token is automatically generated at startup (valid only on the local machine):

cyntex server --role=all
#         Bootstrap token: cyxt_xxxxxxxxxxxx

Alpha Toolset (Read-Only + Scaffold)

Tool	Description
`list_sources`	List all registered sources
`list_pipelines`	List all pipelines and their runtime status
`get_pipeline_status`	Query the status, lag, and error rate of a specific pipeline
`get_source_schema`	Get the table schema of a source (DiscoverSchema)
`scaffold_pipeline`	Generate a pipeline YAML draft from a source id and table name
`explain_field`	Look up documentation for any DSL field
`validate_yaml`	Validate a YAML content snippet (three-layer validation)

Beta Toolset (Full CRUD + Lifecycle)

In the Beta phase, all control core operations will be exposed as MCP tools:

create_source / update_source / delete_source
apply_pipeline / delete_pipeline
start_pipeline / stop_pipeline / restart_pipeline
export_pipeline (canonical YAML round-trip)

Using with Claude

Add Cyntex MCP to Claude’s MCP configuration:

{
  "mcpServers": {
    "cyntex": {
      "url": "http://localhost:7778/mcp",
      "headers": {
        "Authorization": "Bearer cyxt_xxxxxxxxxxxx"
      }
    }
  }
}

Once configured, Claude can directly operate your Cyntex instance — create sources, deploy pipelines, and check runtime status.