NewAgent Observability — find failures before customers do.
The vendor directory

AI observability tools: what your agent can actually query

We looked at 91 monitoring and developer tools to see which have an MCP server, the connector that lets your AI coding agent query a tool directly. Each one gets the same four checks: is it official, maintained, easy to install and actually usable? Every fact comes from the vendor's own docs, repos and package registries, with the date we checked it.

Browse by the part of the work you are in

1Code, tests & reviewAgents writing code, generating tests and reviewing pull requests as one loop.2CI & releasePipelines running far more often, and releases compressing from weeks to hours.3Production observabilityInfrastructure, APM, logs, traces, profiling and incident response, with an agent as the reader.4Agent observability & evalsWatching the agents themselves: hallucination, drift, eval scores and what a session actually did.5AI cost managementToken cost, model routing and spend governance.
Filter by:

Official server means the vendor publishes it. Official, archived means the vendor published one and no longer maintains it. Community only means someone outside the vendor built it. MCP client or gateway means the tool connects to other servers; its guide covers any server of its own. No server found means no server exposing the tool's data was found when its guide was checked.

Allure TestOps logo

Allure TestOps

Official server13 tools built into TestOps from 26.1.1. A public beta that may later carry a fee.Code & reviewRead the guideAmazon CloudWatch logo

Amazon CloudWatch

Official server19 read-only tools over Logs Insights, metrics, PromQL and alarms. Runs locally, and log queries are billed by data scanned.ProductionAgent obs & evalsRead the guideAnodot logo

Anodot

Official serverUmbrella, Anodot's former cost unit, hosts an MCP server for AWS, Azure and GCP spend. Anomalies are reached through REST APIs.ProductionNo page planned: effectively dead or absorbed; page decision openRead the guideArize Phoenix logo

Arize Phoenix

Official server27 tools reaching traces and spans, not just evaluations. Works against self-hosted Phoenix too.Code & reviewAgent obs & evalsRead the guideAviator logo

Aviator

Official server3 runbook tools. Aviator's own docs call this the legacy path and prefer its CLI.CI & releaseRead the guideAWS Cost Explorer logo

AWS Cost Explorer

Official server46 tools over billing, budgets and optimization. At 38 tools it measured about 26,390 tokens, the fifth-heaviest here.AI costRead the guideAzure Cost Management logo

Azure Cost Management

Official serverMicrosoft's ARM server reads real Azure spend. 14 cost tools, 5 on by default, GitHub Copilot clients.AI costRead the guideBigPanda logo

BigPanda

Official serverSign in from your AI tool and get ask_biggy, search_knowledge and get_conversation. Action plans become tools on a separate API endpoint.ProductionRead the guideBraintrust logo

Braintrust

Official server42 documented tools across 11 categories, led by SQL over your own logs.Code & reviewAgent obs & evalsRead the guideBuildPulse logo

BuildPulse

Official server9 read-only tools for flaky-test triage and CI failures. The 5-tool 0.1.5 build cost about 686 tokens.Code & reviewRead the guideCircleCI logo

CircleCI

Official server13 tools for pipelines, builds and flaky tests. The npm package is deprecated; use the hosted server.CI & releaseRead the guideClaude Code logo

Claude Code

Official serverMostly an MCP client. `claude mcp serve` exposes its own tools over stdio.Code & reviewRead the guide

Showing 12 of 91

The pattern across all 91 vendors

An MCP server can only return what its backend kept

Connectors cap results and sample telemetry, usually without saying so. An agent can then report that nothing was found when the records exist. Oodle's own MCP server reads telemetry that was never sampled or rolled up, so the agent sees all of it.

See agent observability