Every major observability platform now has an official MCP server. That happened fast — in early 2025, only Sentry had one. By March 2026, Datadog, Grafana, New Relic, Honeycomb, PagerDuty, and Splunk have all shipped their own. The observability category is the most “first-party” MCP space we’ve covered.
This matters because observability is where AI agents add the most obvious value. Debugging production errors, correlating metrics with deploys, querying logs in natural language — these are tasks where context switching between your IDE and a dashboard wastes real time. An MCP server that puts your observability data inside your agent’s context eliminates that round-trip.
We’ve now reviewed all six: the Sentry MCP server (4/5), the Grafana MCP server (4/5), the Datadog MCP server (4/5), the New Relic MCP server (4/5), the Honeycomb MCP server (4/5), and the PagerDuty MCP server (4/5). Here’s how the full observability landscape compares, and which ones are worth integrating.
The Contenders
| Server | Type | Stars | Transport | Auth Model | Tools | Free Tier | Best For |
|---|---|---|---|---|---|---|---|
| Sentry | Error tracking | 579 | Remote (SSE) | OAuth 2.0 | ~20 | Yes (10K events/mo) | Debugging production errors |
| Datadog | Full-stack APM | 7 (hosted) | Remote (HTTP) | OAuth / API key | 50+ | No (14-day trial) | Enterprise full-stack observability |
| Grafana | Dashboards + LGTM | 2,500 | stdio + SSE + HTTP | API token | 40+ | Yes (Grafana Cloud free) | Metrics/logs/traces visualization |
| New Relic | Full-stack APM | 3 (hosted) | Remote (HTTP) | API key / OAuth | 35 | Yes (100GB/mo) | Natural language observability queries |
| Honeycomb | Event-based observability | 41 (deprecated repo) | Remote (hosted) | OAuth 2.1 / API key | 14+ | Yes (20M events/mo) | High-cardinality event analysis |
| PagerDuty | Incident management | 56 | Both (hosted + local) | API token | 67 | Yes (5 users) | Incident response & on-call |
Notable addition: Splunk has an MCP server for its Cloud Platform (updated Feb 2026), but it’s primarily an enterprise play and not widely adopted in the MCP ecosystem yet. We’re tracking it but not comparing it in detail here.
Three Observability Layers
Observability MCP servers cover different layers of the stack, and understanding which layer you need matters more than comparing feature counts:
1. Error Tracking & Debugging (Sentry, Honeycomb)
These servers focus on what went wrong. Sentry is purpose-built for errors — stack traces, issue grouping, release tracking, session replays. Honeycomb takes a broader event-based approach with high-cardinality tracing, letting you slice errors by any dimension.
When to use this layer: You’re a developer debugging a production error. You need the stack trace, the user who triggered it, and the deploy that introduced it. Your agent investigates the specific issue.
2. Full-Stack APM & Metrics (Datadog, New Relic, Grafana)
These servers cover everything — metrics, traces, logs, infrastructure, APM, and more. They’re the “one platform for all telemetry” approach. Datadog’s MCP server alone has toolsets for core monitoring, alerting, APM, database monitoring, error tracking, feature flags, LLM observability, product analytics, and synthetic testing.
When to use this layer: You’re investigating a performance degradation, correlating a metric spike with a deploy, or querying across infrastructure and application data. Your agent needs the big picture.
3. Incident Management (PagerDuty)
PagerDuty doesn’t collect telemetry — it manages the human response to incidents. Creating incidents, acknowledging alerts, escalating to the right on-call engineer, coordinating resolution. It complements the other tools rather than replacing them.
When to use this layer: You’re on-call and an alert fires. Your agent acknowledges the page, pulls context from Sentry/Datadog, and helps you coordinate the response without switching between five dashboards.
The Servers in Detail
Sentry MCP — The Error Tracking Specialist
getsentry/sentry-mcp | Our full review | Rating: 4/5
Sentry’s MCP server sets the standard for how first-party integrations should work. The killer feature combination: OAuth 2.0 authentication (no API tokens on disk) + zero-install remote hosting at mcp.sentry.dev + Seer AI integration for automated root cause analysis.
The ~20 tools cover issue investigation, event analysis, natural language search, and project management. The Seer integration is the genuine differentiator — when you call get_issue_analysis, Sentry’s proprietary AI doesn’t just return the stack trace, it explains why the error happened and suggests fixes. No community server can replicate this.
Strengths: Best auth model in any MCP server (OAuth 2.0), zero install, Seer AI analysis, deep error-tracking tools, first-party maintenance.
Weaknesses: 748+ GitHub issues at pre-1.0 (v0.29.0), cross-project queries fail, AI search needs a separate LLM key, 63 npm dependencies.
Best for: Developers already on Sentry Cloud who debug production errors from their IDE.
Datadog MCP — The Enterprise Swiss Army Knife
Datadog MCP Server | Our full review | Rating: 4/5
Datadog’s MCP server is the most feature-rich observability integration in the MCP ecosystem. It’s built around toolsets — modular capability groups you enable or disable via URL parameters to save context window space. Available toolsets: core (logs, metrics, traces, dashboards, monitors, incidents, hosts, services, events, notebooks), alerting, APM, database monitoring, error tracking, feature flags, LLM observability, product analytics, networks, security, software delivery, and synthetics.
This is the only MCP server that covers LLM observability (monitoring your AI agents’ performance), feature flag management, database monitoring, and synthetic testing alongside traditional APM. For teams already on Datadog, this puts an enormous amount of operational context at your agent’s fingertips.
The agent-native design is what sets it apart from API wrappers. Token-budget pagination prevents huge logs from consuming the context window. CSV output uses 50% fewer tokens than JSON for tabular data. SQL-like log queries (SELECT service, COUNT(*) FROM logs WHERE status='error' GROUP BY service) achieve 40% cost reduction compared to raw retrieval. Error messages include suggestions (“unknown field ‘stauts’ — did you mean ‘status’?").
The server connects via Streamable HTTP to managed regional endpoints (US1, US3, EU1, AP1/AP2) — zero install required. Auth supports OAuth 2.0 for interactive flows and API key + application key headers for headless setups. Works with Claude Code, Cursor, OpenAI Codex, GitHub Copilot, VS Code, Goose, Cognition, and Kiro.
Strengths: Broadest tool coverage (50+ tools across 10+ toolsets), agent-native design (token-budget pagination, CSV formatting, SQL-like queries), modular toolsets via URL parameters, LLM observability (unique), zero-install remote hosting with regional endpoints, RBAC-aware security with HIPAA support, enterprise-ready, GA status.
Weaknesses: API key auth by default (OAuth available but not primary), Datadog’s pricing is notoriously complex (no permanent free tier), closed-source (can’t audit or self-host), /api/unstable/ path despite GA status, not GovCloud compatible, incident timeline data missing, community server (winor30/mcp-server-datadog, 139 stars) covers gaps the official server doesn’t (host muting, downtimes, RUM).
Best for: Enterprise teams already on Datadog who want comprehensive operational context in their coding agents.
Grafana MCP — The Open-Source Metrics Gateway
grafana/mcp-grafana | Our full review | Rating: 4/5
Grafana’s MCP server is the most popular observability MCP server by GitHub stars (2,500), and the only one with a truly open-source architecture. It connects to your Grafana instance (9.0+) and the surrounding LGTM stack — Loki for logs, Tempo for traces, Mimir for metrics, and Grafana’s own dashboards.
The 40+ tools span 15 configurable categories: dashboard management, Prometheus, Loki, ClickHouse, CloudWatch, Elasticsearch, log search, incidents, Sift investigations, alerting, OnCall, navigation, annotations, rendering, and admin. The --disable-<category> and --enabled-tools flags let you trim this to exactly the tools you need — the most granular context window management of any MCP server we’ve reviewed. Grafana also ships separate dedicated MCP servers for Loki (log querying) and Tempo (distributed tracing), so you can go deep on specific telemetry types.
A key differentiator: Grafana’s MCP server works with any backend data source Grafana supports — Prometheus, InfluxDB, Elasticsearch, CloudWatch, and dozens more. You’re not locked into one vendor’s telemetry pipeline. The incident management loop (alerting → OnCall → Sift investigations → dashboards) is the most complete in any single MCP server.
Strengths: Open source (Apache 2.0), largest GitHub community (2,500 stars), works with any Grafana data source, configurable tool categories, separate Loki/Tempo servers for deep dives, self-hostable, all three transports (stdio + SSE + Streamable HTTP), active development (15+ releases in 4 months), agent-aware dashboard tools (get_dashboard_summary, JSONPath extraction, patch_dashboard), panel/dashboard PNG rendering.
Weaknesses: No hosted remote server (must run yourself), service account token auth (not OAuth), 61 open issues including security findings (#608 — TLS bypass, credential exposure), requires Grafana 9.0+, 16K token tool description footprint, some tool categories require Grafana Cloud (incidents, OnCall, Sift).
Best for: Teams running their own Grafana stack who want agent-assisted metrics, logs, traces, alerting, and incident management queries.
New Relic MCP — Natural Language Observability
New Relic AI MCP Server (4/5) | Docs | Public Preview
New Relic’s MCP server leans heavily into natural language querying. Its standout feature: you ask questions in plain English (“show me the slowest transactions in the payments service this week”), and the server translates them into NRQL (New Relic Query Language) queries via the natural_language_to_nrql_query tool. You don’t need to learn NRQL to query your observability data.
The server provides 35 tools across 6 categories: discovery (entities, accounts), data access (NRQL, logs), alerting (policies, conditions), incident response (issues, deployments, error groups), performance analytics (golden metrics, transactions), and advanced analysis (alert reports, user impact reports, synthetics). It’s hosted at mcp.newrelic.com with Streamable HTTP transport — zero-install, always current. Auth supports both API keys and OAuth 2.0.
New Relic’s generous free tier (100GB/mo ingestion, no credit card required) makes it the most accessible full-stack observability MCP server. The analyze_golden_metrics tool surfaces throughput, response time, error rate, and saturation as a first-class tool — most other servers force agents to derive these from raw queries. The analyze_deployment_impact tool correlates deploys with performance changes automatically.
The tag-based tool filtering system (include-tags headers) lets you present only relevant tool categories per request — cleaner than URL parameters (Datadog) or CLI flags (Grafana). But with 35 tools it trails Datadog (50+) and Grafana (40+), and the tool set is read-only — no muting, no downtime scheduling, no alert acknowledgment. Six community alternatives (cloudbring/newrelic-mcp with 18 tools being the most complete) cover some of these gaps.
Strengths: Natural language to NRQL translation (no query language learning curve), 35 tools across 6 categories, best free tier in category (100GB/mo), golden metrics as a dedicated tool, deployment impact analysis, tag-based tool filtering, remote hosting, RBAC-aware security.
Weaknesses: Public Preview (not GA), no FedRAMP/HIPAA support, API key auth by default, minimal GitHub presence (3 stars, 2 commits), read-only orientation (no write operations), 6 community alternatives suggest gaps in the official server, trails Datadog and Grafana in tool count.
Best for: Teams on New Relic who want natural language observability querying with the lowest barrier to entry in the category.
Honeycomb MCP — High-Cardinality Event Analysis
honeycombio/honeycomb-mcp | Our full review | Rating: 4/5
Honeycomb is the pioneer of observability-as-debugging — treating every request as a structured event with arbitrary dimensions, then letting you slice and dice without pre-defined dashboards. Their MCP integration has an unusual history: they built an open-source self-hosted server (41 stars, MIT, 14 tools), then deprecated it in favor of a hosted MCP server at mcp.honeycomb.io/mcp.
The self-hosted server exposed 14 well-documented tools: run_query (analytics with calculations, breakdowns, filters), analyze_columns (statistical analysis), list_datasets/get_columns (schema exploration), list_slos/get_slo, list_triggers/get_trigger, list_boards/get_board, list_markers, list_recipients, get_trace_link (deep links), and get_instrumentation_help (OpenTelemetry guidance). The hosted server continues these capabilities with BubbleUp anomaly analysis — Honeycomb’s signature feature that automatically identifies what’s different about a subset of events compared to the baseline.
The hosted server uses OAuth 2.1 for interactive clients (browser-based login, team selection, scope grants) and Management API keys for headless agents. It requires the mcp-remote npm bridge for stdio clients. Multi-region support with separate US and EU endpoints addresses data residency. Works on all Honeycomb tiers including Free (20M events/mo) — a major improvement from the deprecated server which required Enterprise.
Strengths: OAuth 2.1 authentication (matches Sentry as best auth in category), BubbleUp anomaly decomposition (unique — no other MCP server has automated anomaly analysis), high-cardinality event debugging by design, OpenTelemetry instrumentation guidance, available on all tiers including free (20M events/mo), multi-region (US/EU endpoints), zero-install hosted server.
Weaknesses: Messy deprecation transition (self-hosted deprecated, hosted tools not publicly documented at same granularity), 15 open issues on deprecated repo including hallucinated columns (#24), tight rate limits (50 calls/min, 10/min for service map), 24-hour session timeouts, can’t add queries to existing boards, mcp-remote bridge dependency for stdio clients, fewer tools than Datadog (50+) or Grafana (40+), Enterprise features gated (service map, SLO reporting).
Best for: Teams already on Honeycomb who do high-cardinality debugging on distributed systems and want OAuth-first agent integration with BubbleUp analysis.
PagerDuty MCP — Incident Response Automation
PagerDuty/pagerduty-mcp-server | Our full review | Rating: 4/5
PagerDuty is the outlier in this comparison — it doesn’t collect telemetry. Instead, it manages the human response to incidents: creating/acknowledging/resolving incidents, managing on-call schedules, escalating to the right engineer, and coordinating across teams.
PagerDuty offers both a hosted MCP service at mcp.pagerduty.com/mcp and a self-hosted open-source server (Apache-2.0, Python). This dual-deployment approach gives teams flexibility — use the hosted server for convenience or self-host for compliance requirements.
67 tools across 13 categories — the largest tool count of any server in this comparison. 14 incident tools (create, resolve, merge, snooze, response plays, multi-responder coordination), 8 event orchestration tools, 7 status page tools, 7 team tools, 6 schedule tools, 5 alert grouping tools, 4 change event tools, 4 service tools, 3 workflow tools, 2 escalation policy tools, 2 user tools, 2 log entry tools, 1 on-call tool. Read-only by default — 31 tools are enabled by default, the remaining 36 write tools require --enable-write-tools. This is the safest write-access model of any MCP server we’ve reviewed.
The real value is combining PagerDuty MCP with another observability MCP server: your agent gets paged, acknowledges the incident via PagerDuty, pulls the error details from Sentry, checks the metrics in Grafana, and writes up the incident summary — all without you switching tabs.
Strengths: 67 tools (most comprehensive incident MCP), read-only defaults with explicit write opt-in, dual deployment (hosted + self-hosted), event orchestration (unique — no other server lets agents configure routing rules), status page management, Docker support, Apache-2.0 license, active community (267 commits, 27 forks).
Weaknesses: API token auth only (no OAuth browser flow), no HTTP transport for self-hosted (issue #25), corporate proxy issues (issue #66), pagination bugs (issue #62), no AI analysis layer (pure API wrapper), limited free tier (5 users), 67 tools exceeds their own 20-25 recommendation.
Best for: On-call engineers who want their AI agent to help manage incident response alongside debugging.
Feature Comparison
| Feature | Sentry | Datadog | Grafana | New Relic | Honeycomb | PagerDuty |
|---|---|---|---|---|---|---|
| Error tracking | Deep | Yes | Via data source | Yes | Events | No |
| APM/traces | No | Deep | Via Tempo | Yes | Deep | No |
| Metrics | No | Deep | Deep (any source) | Yes | Limited | No |
| Logs | No | Deep | Via Loki | Yes | Yes | No |
| Infrastructure | No | Deep | Via data source | Yes | No | No |
| Incident mgmt | No | Alerting | Alerting | Alerting | Triggers | Deep |
| AI analysis | Seer (proprietary) | Bits AI | No | NRQL translation | Natural language | No |
| Auth model | OAuth 2.0 | OAuth / API key | API token | API key / OAuth | OAuth 2.1 / API key | OAuth 2.0 |
| Transport | Remote (SSE) | Remote (HTTP) | stdio + SSE + HTTP | Remote (HTTP) | Remote (HTTP) | Both |
| Open source | Yes | No | Yes | No (hosted) | Deprecated (was MIT) | Yes |
| Status | Pre-1.0 | GA | Active | Public Preview | Hosted (self-hosted deprecated) | Active |
| Rating | 4/5 | 4/5 | 4/5 | 4/5 | 4/5 | 4/5 |
How to Choose
Start with what you already use. This is the most important rule in observability MCP servers. Every server in this comparison is tightly coupled to its platform — you need to already be sending telemetry to that platform for the MCP server to be useful. There’s no “best observability MCP server” in isolation.
Decision Flowchart
“I debug production errors daily” → Sentry MCP (4/5). Deepest error investigation tools, OAuth auth, Seer AI analysis. Pair with PagerDuty if you’re on-call.
“I need the full picture — metrics, traces, logs, everything” → Datadog MCP (4/5) if you can afford it (broadest toolset, 10+ toolsets, GA, agent-native design). New Relic MCP (4/5) if you want a generous free tier, natural language querying, and golden metrics analysis.
“I run my own observability stack” → Grafana MCP (4/5). The only server that works with any backend data source. Open source, self-hostable, configurable tool categories, full incident pipeline.
“I need to debug complex distributed systems” → Honeycomb MCP (4/5). High-cardinality event analysis with BubbleUp anomaly decomposition is purpose-built for finding needles in haystacks. OAuth 2.1, hosted, works on all tiers including Free.
“I’m on-call and want to manage incidents from my agent” → PagerDuty MCP (4/5). 67 tools with read-only defaults. Pair it with whichever debugging server you use.
The Stack We’d Recommend
For most teams, you want two observability MCP servers:
- A debugging server — Sentry (error-focused), Datadog or New Relic (full-stack), or Honeycomb (event-based)
- An incident server — PagerDuty (if you use it for on-call)
Add Grafana if you have custom dashboards you want your agent to query, or if you run a multi-vendor observability stack.
Don’t try to install all six. MCP servers compete for context window space, and each one you add reduces the tokens available for actual work. Pick the ones that match your existing observability stack.
The Bigger Picture
Observability is the most “first-party” category in the MCP ecosystem. Every major platform — Sentry, Datadog, Grafana, New Relic, Honeycomb, PagerDuty, and Splunk — has shipped an official MCP server. Compare this to databases (where both official MCP servers are archived) or search (where Google has no MCP server at all).
This makes sense: observability data is exactly the kind of context that makes AI agents more useful. Telling your agent “here’s the error, here’s the metrics, here’s who’s on-call” is better than alt-tabbing between six dashboards. The platforms know this, and they’re racing to be the ones feeding that context.
The trend we’re watching: Datadog and Sentry lead on design innovation — Sentry with OAuth 2.0 auth, Datadog (4/5) with agent-native tool design (token-budget pagination, CSV formatting, SQL-like queries, modular toolsets). If other categories adopt these patterns, MCP servers become dramatically easier to set up and more efficient to use.
For more on each category, see our full comparison guides: browser automation, databases, web scraping, memory, search, documentation, productivity, and DevOps & infrastructure.