Datadog MCP Server

Yes, conditionally: start with production investigation, not unattended remediation. The strongest fit is an SRE, platform engineer or developer already using Datadog who wants an agent to connect evidence across telemetry products.

The main advantage is first-party coverage with inherited permissions, configurable tools and attributable activity. The limitation is operational authority: the surface includes writes, the hosted backend is not independently inspectable in the public repository, and correctness still depends on filters, time windows and returned data.

Poor fit: air-gapped deployments, self-host-only requirements, and unrestricted autonomous changes to production resources. The 77/100 score is an editorial judgment from this evidence review, not a controlled benchmark or vendor rating. Read our methodology.

• Investigate incidents: retrieve logs, metrics and traces around a known service and time window.
• Diagnose application performance: combine APM and profiling evidence to narrow a hypothesis.
• Inspect operational state: review incidents, monitors, alerts and service health from an MCP client.
• Manage resources: selected toolsets support changes to dashboards, monitors and workflows when permission allows.
• Run advanced investigations: the code-execution toolset supports sandboxed TypeScript queries against Datadog APIs.
• Support security reviews: use the relevant security and audit tools within product permissions.

Choose toolsets for the job instead of exposing the full catalog. The tool inventory changes frequently; this profile does not assert a fixed tool count.

Use OAuth where supported and a dedicated, least-privilege service account if keys are necessary. Datadog documents separate MCP read/write access alongside product permissions. Results returned to the connected client can become part of that client’s model context; review that data boundary before using sensitive telemetry.

Concrete write-control evidence: Datadog’s September 22 changelog says saved-query creation and updates previously bypassed the organization’s MCP write-protection setting. The entry records a fix, not proof that the defect remains open. It supports a practical recommendation: test a denied write before granting operational access. Read the fix ↗

Auditing is disclosed. Datadog records MCP tool activity and emits usage metrics. The evaluation rubric assigns an F under its strict “any telemetry” scan rule; that label describes the rule, not a malware finding. We treat auditability as a useful control while making the logging boundary explicit.

The managed sandbox is a separate capability from local shell access. Do not assume enabling code execution authorizes arbitrary network access or production writes. Review Datadog’s current data-handling documentation ↗

The overall assessment emphasizes safety, retrieval correctness and operational reliability rather than averaging every score. Scores are editorial assessments, not measured success rates.

Protocol compatibility: Streamable HTTP is documented. Specific support for the 2026-07-28 protocol was not independently verified in this review; transport choice alone does not establish a protocol version.

Estimated evaluation usage: 62,000 tokens — 50,000 input and 12,000 output. This budgets for a multi-step evaluation covering metrics, logs, trace correlation, access restrictions, a controlled write check and audit review. It is not the cost of each request or a measured tool-schema size.

Use toolsets, omit_tools, narrow queries and supported max_tokens settings to reduce context consumption. Start with the smallest useful toolset; add APM or other capabilities only when needed.

Estimated API cost for that workload

The examples below are rate scenarios, not current quotes for named models. Apply your provider’s actual per-million-token rates.

Cost = 0.05 × input rate + 0.012 × output rate. Figures exclude caching, retries, subscriptions and Datadog charges.

Datadog-side budget: account for underlying product subscriptions, applicable Audit Trail costs and configuration/review time. Current documented fair-use limits are 50 requests per 10 seconds and 100,000 monthly tool calls; limits can change. Check current limits ↗

1. Confirm your Datadog region and create or select a read-only role with the needed product access.
2. Add the official remote endpoint to your client and complete OAuth.
3. Limit available toolsets and run a query against known, non-sensitive telemetry.
4. Compare returned filters, time windows and results with the Datadog UI.
5. Check that a deliberately unavailable resource remains denied and that calls appear in Audit Trail.
6. Enable writes only for a defined workflow, with human approval and a controlled test.

Claude Code — US1 example

claude mcp add --transport http datadog https://mcp.datadoghq.com/v1/mcp

For EU, use https://mcp.datadoghq.eu/v1/mcp. Other sites and clients have their own configuration details. Follow the official setup instructions ↗

Compare available MCP clients and browse setup guides. Keep secrets in the client’s credential mechanism, never in prompts or shared examples.

These linked first-party records support the assessment. Maintenance activity is encouraging, but a vendor changelog is not an independent production benchmark.

September 22, 2026

Saved-query writes now honor the organization’s MCP write-protection setting. Official record ↗

September 23, 2026

Spreadsheet pagination was corrected after boundary errors made some results unreachable. Official record ↗

September 8, 2026

Schema fixes restored tool loading in Vertex AI and Gemini; database schema results gained previously empty fields. Official record ↗

September 9, 2026

Network availability-zone filtering was corrected. Official record ↗

September 3, 2026

Network interpretation and oversized code-execution result handling were improved. Official record ↗

August 28, 2026

The documented monthly default quota increased to 100,000 calls. Official record ↗

Evidence limits: this review does not include a controlled uptime test, incident-resolution benchmark or measured schema-token capture. Confidence remains moderate because the strongest recent evidence is first-party.

Official repository ↗ · Documentation and data handling ↗