Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.agentmark.co/llms.txt

Use this file to discover all available pages before exploring further.

​
MCP Trace Server

The @agentmark-ai/mcp-server package exposes your AgentMark traces to AI-powered editors through the Model Context Protocol (MCP). This enables AI assistants to query and debug your agent executions directly.
This is different from the AI Editor Integration which helps AI editors create AgentMark files. The MCP Trace Server provides trace debugging capabilities - allowing AI assistants to inspect execution logs, find errors, and analyze performance.

​
Overview

The MCP Trace Server provides two tools:
ToolDescription
list_tracesList recent traces with metadata (status, latency, cost, tokens)
get_traceGet detailed trace data with span filtering and pagination

​
Installation

npm install @agentmark-ai/mcp-server

​
Configuration

​
Environment Variables

VariableDefaultDescription
AGENTMARK_URLhttp://localhost:9418URL of the AgentMark API server
AGENTMARK_API_KEY-API key for authentication (required for cloud)
AGENTMARK_TIMEOUT_MS30000Request timeout in milliseconds

​
Editor Setup

Add to your project’s .mcp.json:
{
  "mcpServers": {
    "agentmark-traces": {
      "command": "npx",
      "args": ["@agentmark-ai/mcp-server"],
      "env": {
        "AGENTMARK_URL": "http://localhost:9418"
      }
    }
  }
}

​
Cloud Configuration

For AgentMark Cloud integration, add your API key: 
{
  "mcpServers": {
    "agentmark-traces": {
      "command": "npx",
      "args": ["@agentmark-ai/mcp-server"],
      "env": {
        "AGENTMARK_URL": "https://api.agentmark.co",
        "AGENTMARK_API_KEY": "your-api-key"
      }
    }
  }
}

​
Tools Reference

​
list_traces

List recent traces with metadata. Returns trace IDs, names, status, latency, cost, and token counts. Parameters:
ParameterTypeDefaultDescription
limitnumber50Maximum traces to return (max: 200)
sessionIdstring-Filter traces by session ID
datasetRunIdstring-Filter traces by dataset run ID
cursorstring-Pagination cursor from previous response
Response: 
{
  "items": [
    {
      "id": "trace-abc123",
      "name": "my-prompt",
      "status": "0",
      "latency": 1234,
      "cost": 0.05,
      "tokens": 500,
      "start": 1704067200000,
      "end": 1704067201234
    }
  ],
  "cursor": "eyJvZmZzZXQiOjUwfQ==",
  "hasMore": true
}
Status Values:
  • 0 = OK
  • 1 = Warning
  • 2 = Error

​
get_trace

Get trace summary with filtered and paginated spans. Use this to drill into specific trace details. Parameters:
ParameterTypeDefaultDescription
traceIdstring-Required. The trace ID to retrieve
filtersarray-Filter criteria for spans (see Filter Schema)
limitnumber50Results per page (max: 200)
cursorstring-Pagination cursor from previous response
Filter Schema: Each filter object has three fields:
FieldTypeDescription
fieldstringField to filter on
operatorstringComparison operator
valuestring | numberValue to compare against
Supported Fields and Operators:
FieldOperatorsDescription
statuseqSpan status (“0”=ok, “1”=warning, “2”=error)
durationgt, gte, lt, lteSpan duration in milliseconds
namecontainsSpan name substring match
data.typeeqSpan type (“GENERATION”, “SPAN”, “EVENT”)
data.modelcontainsModel name substring match
Example - Find error spans: 
{
  "traceId": "trace-abc123",
  "filters": [
    { "field": "status", "operator": "eq", "value": "2" }
  ]
}
Example - Find slow LLM generations: 
{
  "traceId": "trace-abc123",
  "filters": [
    { "field": "data.type", "operator": "eq", "value": "GENERATION" },
    { "field": "duration", "operator": "gt", "value": 5000 }
  ]
}
Response: 
{
  "trace": {
    "id": "trace-abc123",
    "name": "my-prompt",
    "spans": [],
    "data": {
      "status": "0",
      "latency": 1234,
      "cost": 0.05,
      "tokens": 500
    }
  },
  "spans": {
    "items": [
      {
        "id": "span-xyz789",
        "name": "llm-call",
        "duration": 892,
        "parentId": null,
        "timestamp": 1704067200000,
        "traceId": "trace-abc123",
        "status": "0",
        "data": {
          "type": "GENERATION",
          "model": "gpt-4o-mini",
          "inputTokens": 150,
          "outputTokens": 350,
          "cost": 0.05
        }
      }
    ],
    "cursor": "eyJvZmZzZXQiOjUwfQ==",
    "hasMore": false
  }
}

​
Error Handling

All tools return structured errors with error codes: 
{
  "error": "Trace not found: trace-abc123",
  "code": "NOT_FOUND",
  "details": { "traceId": "trace-abc123" }
}
Error Codes:
CodeDescription
CONNECTION_FAILEDCannot reach the AgentMark server
INVALID_QUERYMalformed filter or unsupported field/operator
NOT_FOUNDTrace or resource does not exist
TIMEOUTRequest exceeded time limit

​
Requirements

For local development, the MCP server connects to the AgentMark CLI local API server:
  1. Run agentmark dev to start the local development server
  2. Execute prompts to generate traces
  3. Use your AI editor to query and debug traces

​
Programmatic Usage

You can also use the server programmatically: 
import { createMCPServer, runServer } from '@agentmark-ai/mcp-server';

// Run with stdio transport (for MCP clients)
await runServer();

// Or create a server instance for custom transport
const server = await createMCPServer();

​
Related Documentation

AI Editor Integration

Help AI editors create AgentMark files

Traces and Logs

Learn about AgentMark tracing

CLI Reference

Start the dev server with agentmark dev

​
Have Questions?

We’re here to help! Choose the best way to reach us: