MCP Server
@monitor/mcp transforms the monitor server's HTTP API into an agent-friendly MCP tool surface. It can be used as a supplementary method for automatic collection paths like Claude plugins, and in manual client environments, it serves as a core adapter.
Core Files
packages/mcp/src/index.tspackages/mcp/src/client.tspackages/mcp/src/result.tspackages/mcp/test/client.test.ts
Current Number of Tools
Currently createMonitorMcpServer() registers 24 monitor_* tools. These tools are divided into lifecycle, event logging, semantic flow, and workflow library.
How It Works
- MCP client invokes a tool.
- The handler in
packages/mcp/src/index.tsvalidates input schema. MonitorClientsends an HTTP request to the corresponding/api/*endpoint.- The result is wrapped in MCP response format via
toToolResponse().
Why This Layer Is Important
Standardizes manual observability paths
Not all agent runtimes provide hooks or plugins. With an MCP surface, manual monitoring can at least target the same HTTP server.
Prevents failures from cascading to the entire agent task
MonitorClient does not throw exceptions for network errors or abnormal HTTP responses; instead, it returns ok: false results. In other words, even if the monitor server goes down, the agent task itself continues.
Connects to runtime adapter documentation
Many monitor_* calls mentioned in Claude plugin documentation and manual integration documentation actually reside in this package.
Characteristics of the Current Structure
- All tools are manually registered in one file.
- Tool name, input schema, and target endpoint are close together, making it easy to read.
- However, as the number of tools grows, repetitive code and drift risk also increase.
Maintenance Checkpoints
- When creating a new HTTP endpoint, verify if MCP surface is also needed
- Verify that tool input schema and server request schema have the same constraints
- Verify that tool names in runtime documentation match the actual registered names
- For features with canonical paths like
monitor_user_message,monitor_assistant_response, and workflow library tools, converge to a single path in documentation and skills