# Agentic Observability MCP [Health: Active]

**Category:** 📊 Monitoring  
**Repository:** https://github.com/mdfifty50-boop/agent-observability-mcp  
**GitHub Stars:** 0  
**npm Downloads (last month):** 168  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/agentic-observability-mcp

## Description
Agent tracing, cost tracking, anomaly detection for LLM agents

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "agentic-observability-mcp": {
    "command": "npx",
    "args": ["agentic-observability-mcp"]
  }
}
```

## Documentation

## What Agentic Observability MCP MCP server does

Agentic Observability MCP MCP server provides MCP tools for collecting operational data from AI agent runs. It can record general actions such as decisions, model requests, tool calls, and errors; track input and output tokens; calculate call and session costs; and summarize activity across a session.

The server also evaluates recorded activity for several predefined conditions. These include unusually high costs, elevated tool failure rates, long-running tool calls, repeated calls with identical parameters, and excessive token use. Decision traces can serve as audit records when their descriptions include the relevant reasoning or context.

## How it works

An agent sends events to the server as it works. Use `trace_agent_action` for broad action records, `track_token_usage` for LLM usage, and `log_tool_call` for MCP tool performance data. `get_cost_report` aggregates spending by dimensions such as model, provider, tool, or session. `get_session_summary` combines costs, tokens, tool statistics, errors, model details, and duration into a session report.

`detect_anomaly` checks stored activity for the supported anomaly categories. Loop detection identifies the same tool and parameters being used at least three times. The documented thresholds include a tool error rate above 30 percent and tool latency above 10 seconds; cost and token checks use configured thresholds.

Two static resources are available: `observability://pricing` exposes the pricing data used for calculations, and `observability://best-practices` provides an observability guidance document.

## Setup and configuration

Agentic Observability MCP MCP server requires Node.js 18 or newer and runs without API keys. Install it through an MCP client by launching `npx agentic-observability-mcp` as a stdio server. Claude Desktop and Claude Code use `~/.claude/settings.json` or a project `.mcp.json`; Cursor uses `.cursor/mcp.json`. Windsurf and VS Code use their MCP configuration files with the same command and argument pattern.

The server has no documented environment variables or external service configuration requirements. It includes pricing entries for selected Anthropic, OpenAI, Google, and Mistral models, as well as zero-cost self-hosted models. Custom pricing can be supplied when the built-in entries do not match the model being tracked.

## Tools and capabilities

- Record agent actions with metadata and timestamps.
- Calculate LLM call costs from token counts and model pricing.
- Group costs by model, provider, tool, or session.
- Measure tool-call latency, success, failure, and error details.
- Produce session-level reports and run anomaly checks.
- Expose pricing and observability guidance as MCP resources.

## Limitations and notes

The first version uses in-memory Maps for storage. Restarting the server removes collected traces, costs, and session data. The repository describes a future storage layer that could be replaced with Redis or Postgres, but those backends are not part of the documented v1 setup.

The pricing table covers named model families rather than every possible provider or model. Cost accuracy therefore depends on selecting the correct model or providing custom pricing.

_Full upstream README: https://allmcps.com/mcp/agentic-observability-mcp/readme_