What is the Datadog MCP server and how do you set it up for a team?

Short answer

The Datadog MCP server is a remote MCP server run by Datadog, generally available since March 9, 2026. It lets AI agents such as Claude Code, Cursor, Codex and Copilot query logs, metrics, traces, monitors, dashboards and more. It uses the signed-in user's own permissions, needs MCP Access switched on by an admin, and only exposes the core toolset unless you ask for more.

What you get

The server gives an AI client remote access to most Datadog products: APM, logs, metrics, monitors, dashboards, security signals and more. A developer can ask “why did checkout latency spike at 14:00?” from Claude Code or Cursor, and the agent searches logs and traces itself instead of the developer copying screenshots into the chat.

Datadog names many clients on its setup page, including Cursor, Claude Code, VS Code, Codex, ChatGPT, Claude, Copilot CLI, Kiro, Goose, OpenCode and Warp. Some tools use AI models hosted by Datadog’s AI providers, for example to build a query from a natural-language question.

Toolsets: start narrow

Tools are grouped in toolsets, chosen with the toolsets query parameter on the endpoint (Codex uses the X-Datadog-MCP-Toolsets header instead).

  1. No parameter: only the default core toolset.
  2. Add what a role needs, for example ?toolsets=apm,llmobs for an engineer debugging services. There are 37 generally available toolsets, among them alerting, apm, dashboards, dbm, kubernetes, logs-configuration, rum, security, synthetics and workflows.
  3. toolsets=all turns on all GA toolsets except chat. Preview toolsets such as live-debugger must be named explicitly.
  4. Remove single tools with omit_tools, for example the tools that create or edit notebooks.

Every tool you add costs context in each request, so a narrow set is also a faster and cheaper agent.

Permissions and admin switches

Layer What it controls
Organization setting MCP Access Whether the MCP server can be used at all. Off means off for everyone.
Organization setting MCP Write Access Whether write tools are visible. Without it, they are hidden from all users.
Role permissions mcp_read and mcp_write Included in the Datadog Standard Role by default; custom roles need them added.
The user’s normal permissions The server forwards the user’s own credentials and cannot grant more. A write such as editing a monitor also needs monitors_write.
RBAC, Data Access Control, log restriction queries Narrow which data an agent sees, the same way they do for the person.
IP allowlist Optional: restrict which networks may connect.

Sign-in is by OAuth 2.0 (recommended), a personal or service access token, or API and application keys sent as headers. For automation, Datadog advises a scoped service account with only the permissions it needs. Every tool call lands in Audit Trail under “MCP Server”.

The common misunderstanding

“The agent will find the root cause.” It will find data fast. What it lacks is the context engineers carry in their heads: which alerts are known noise, which service had the same symptom last quarter and why, which dashboards the team trusts. Datadog’s own use cases show the pattern: in their onboarding example, an agent suggests dashboards from teams that are marked as best-practice examples, which only works because someone put that judgment somewhere the agent can read it.

What this means for a team

Each engineer’s agent now reaches the same telemetry. But what one engineer’s agent learns during an incident (this spike is the nightly batch, ignore that monitor, the fix was a config flag) stays in that one session. The next person, or the next agent, starts again. See MCP servers in Claude Code for a team for sharing the server configuration itself, and how to make AI agents learn from corrections for keeping what was learned.

Where Bombos fits

The Datadog MCP server gives every agent the same view of your systems; Bombos gives them the same memory of how your team reads that view. It is a plugin underneath Claude Code, Codex, Cursor, Copilot and the other AI tools your team uses that learns how your team works, such as “this alert is the nightly batch”, draws on repos and docs, and shows with every answer what it rests on. It runs in the cloud in Europe or standalone on your own server, and one engineer can try it first with their own agents.

Related questions

What is the endpoint of the Datadog MCP server?

It depends on your Datadog site: https://mcp.datadoghq.com/v1/mcp for US1, https://mcp.datadoghq.eu/v1/mcp for EU1, and similar for US3, US5, AP1, AP2 and UK1. The US1-FED and US2-FED government sites are not supported.

How do I add it to Claude Code?

Datadog recommends the plugin: /plugin install datadog@claude-plugins-official. Otherwise use claude mcp add --transport http datadog-mcp with your site's endpoint.

Is there a limit on tool calls?

Datadog applies a fair-use burst limit of 50 tool calls per 10 seconds and no monthly limit. The limit may change; support can raise it.

Does Datadog see my prompts?

No. According to Datadog the server receives individual tool calls (tool name and arguments), not your prompts or the model's reasoning, and it does not send your Datadog data to third-party AI providers.

Sources

  1. Datadog Docs: Datadog MCP Server
  2. Datadog Docs: Set up the Datadog MCP Server
  3. Datadog blog: Datadog MCP Server use cases (March 9, 2026)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.