Error Tracking and Observability MCP Servers Compared (2026): Sentry, Grafana, Honeycomb, New Relic, Axiom, Honeybadger, Dynatrace and PagerDuty

• Sentry: issues, errors, traces and Seer root cause analysis. Hosted at https://mcp.sentry.dev/mcp with OAuth, or local via npx @sentry/mcp-server. The local token needs write scopes. Free: Developer plan, 1 user, 5k errors and 5M spans. Paid: Team from $29/mo.
• Grafana: dashboards, Prometheus, Loki, Tempo, alerting, incidents and OnCall. Hosted Grafana Cloud MCP at https://mcp.grafana.com/mcp with OAuth 2.1, or the open source mcp-grafana server for any Grafana 9.0 or later. Local server has a --disable-write flag. Free: Grafana Cloud Free, 10k active series, 50 GB logs, 50 GB traces, 3 active AI users. Paid: Pro from $19/month plus usage.
• Honeycomb: queries, traces, BubbleUp, service map, boards, triggers and SLOs. Hosted (US and EU), OAuth or API key. 25 read-only tools; 9 write tools need the mcp:write scope. Free: up to 20M events/month, MCP included. Paid: Pro from $150 per month.
• New Relic: NRQL queries, entities, alerts, incidents, errors, logs and distributed traces. Hosted (US, EU, JP), OAuth recommended. The tool reference lists no tools that create or edit alerts or dashboards. Free: 100 GB/month ingest and one full platform user.
• Axiom: APL queries on logs and events, metrics, service map, and full dashboard, monitor and notifier management. Hosted (US), OAuth or token headers. Free: Personal plan, $0, 500 GB/mo data loading compute, 25 GB storage, 10 GB-hours query compute, 30-day retention. Paid: Axiom Cloud from $25/month.
• Honeybadger: errors (faults), Insights logs, uptime, check-ins, deploys and dashboards. Hosted (US and EU) or a Docker image. Self-hosted is read-only unless you allow writes; hosted asks for read-only or read and write access. Free: Developer plan, 5,000 errors/mo, 1 user. Paid: Team from $26/mo.
• Dynatrace: DQL queries, problems, vulnerabilities, Kubernetes events and timeseries forecasting. Hosted per environment, platform token or OAuth client. The local npm server is deprecated. No free plan on the pricing page; 15-day free trial.
• PagerDuty: incidents, alerts, services, schedules, on-call, teams, event orchestrations and status pages. Hosted (US and EU), OAuth client or API token. manage_* tools can create, update and delete, there is no tool filtering, and the docs mention no read-only mode. The self-hosted server is deprecated. Free: 5 users, 1 on-call schedule. Paid: Professional from $21 per user per month billed annually.
• Prices and free tiers were checked on each vendor's pricing page on 2026-10-23 and change often.

Sentry's server lets the agent search issues and events, read issue details and stack traces, and hand an issue to Seer, Sentry's AI root cause tool. The hosted server uses OAuth for every connection. You can scope the URL to one organization or one project, which Sentry recommends.

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/{organizationSlug}/{projectSlug}

{
  "mcpServers": {
    "sentry": {
      "command": "npx",
      "args": ["@sentry/mcp-server@latest"],
      "env": { "SENTRY_ACCESS_TOKEN": "YOUR_SENTRY_TOKEN" }
    }
  }
}

• Main tools named in Sentry's docs: find_organizations, find_projects, search_issues, search_events, get_issue_details, begin_seer_issue_fix.
• Can it change things? Yes. The local token needs the scopes org:read project:read project:write team:read team:write event:write, and Sentry says the server covers triaging issues and managing projects. There is no read-only flag in the README.
• Limit what loads: --disable-skills=seer (or MCP_DISABLE_SKILLS) locally, or ?skills=inspect,triage on the hosted URL.
• Gotcha: search_events and search_issues need an LLM provider on the local server. Set OPENAI_API_KEY (or another provider key) and EMBEDDED_AGENT_PROVIDER. Without one, those tools are missing but the rest work.
• Self-hosted Sentry: add --host=sentry.example.com or set SENTRY_HOST. Seer may not be available there.

• Sentry MCP server: install for every agent

Grafana has two official servers. Grafana Cloud MCP is hosted and uses OAuth 2.1. The open source mcp-grafana server runs locally against any Grafana 9.0 or later, including self-hosted, with a service account token.

claude mcp add grafana --transport http https://mcp.grafana.com/mcp

{
  "mcpServers": {
    "grafana": {
      "command": "uvx",
      "args": ["mcp-grafana", "--disable-write"],
      "env": {
        "GRAFANA_URL": "http://localhost:3000",
        "GRAFANA_SERVICE_ACCOUNT_TOKEN": "YOUR_SERVICE_ACCOUNT_TOKEN"
      }
    }
  }
}

• Main tools (open source): search_dashboards, get_dashboard_by_uid, update_dashboard, query_prometheus, query_loki_logs, find_slow_requests (Tempo, via Sift), list_incidents, create_incident, get_current_oncall_users, alerting_manage_rules.
• Read-only, open source: --disable-write turns off create and update tools, and also hides raw SQL query tools unless you add --enable-query. Turn off whole groups with --disable-<category>, such as --disable-oncall, or pick groups with --enabled-tools.
• Read-only, Grafana Cloud: clear both Write and Query access. Query access allows raw SQL that can modify data. Organization Admins can grant write access; other users only get the Write access option with the Assistant Admin role.
• Gotchas, Grafana Cloud: Grafana Assistant must be available on your stack. Each user who connects through MCP counts as an active user for billing. Only Streamable HTTP is supported, not SSE. For several stacks, use a separate path per stack (https://mcp.grafana.com/mcp/<your-stack>.grafana.net), because Claude Code treats one URL with different X-Grafana-URL headers as a duplicate.
• GRAFANA_API_KEY is deprecated; use GRAFANA_SERVICE_ACCOUNT_TOKEN.

• Grafana MCP server: install for every agent

Honeycomb's hosted server is built for traces and high-cardinality event data. The agent can run queries, open a trace as a waterfall, run BubbleUp to find what is different about slow or failing requests, and read the service map. Honeycomb says to prefer OAuth and use an API key only when the agent cannot do a browser login.

claude mcp add honeycomb --transport http https://mcp.honeycomb.io/mcp

claude mcp add honeycomb --transport http https://mcp.eu1.honeycomb.io/mcp

• Main read tools: run_query, get_query_results, run_bubbleup, get_trace, list_spans, get_service_map, get_dataset_columns, get_slos, get_triggers.
• Write tools: create_board, update_board, create_trigger, update_trigger, create_slo, update_slo, create_recipient, canvas_agent_invoke and feedback. Write tools require the mcp:write scope, so leave it off for read-only use.
• API key setup: send Authorization: Bearer with the key in the form <KEY_ID>:<SECRET_KEY> (stored as HONEYCOMB_API_KEY). Only team owners can make one, and the key should carry no other scopes.
• Limit tools with the X-MCP-Toolset header: all, query or canvas.
• Gotcha: your team must have Honeycomb Intelligence turned on.

• Honeycomb MCP server: install for every agent

New Relic's hosted server lets the agent run NRQL, look up entities and dashboards, read alerts and incidents, and analyze errors, logs, deployments and distributed traces. New Relic says to set up supported clients with OAuth. An admin must turn on MCP access in the feature control manager.

claude mcp add --transport http --scope user newrelic https://mcp.newrelic.com/mcp/
claude mcp login newrelic

• Other regions: https://mcp.eu.newrelic.com/mcp/ (EU) and https://mcp.jp.newrelic.com/mcp/ (JP).
• Main tools: execute_nrql_query, natural_language_to_nrql_query, list_recent_issues, search_incident, list_alert_conditions, search_errors, list_recent_logs, get_distributed_trace_details, analyze_deployment_impact, analyze_golden_metrics.
• Filter tools with the include-tags header, for example discovery,alerting. Tags are discovery, data-access, alerting, incident-response, performance-analytics and advanced-analysis.
• Can it change things? The tool reference lists query, list, search, analyze and report tools only, and the OAuth scope is observability:read. New Relic still warns that AI tools will take actions in your account, so use a least-privilege role.
• Gotchas: Claude Code OAuth needs version 2.1.186 or later. Some tools, such as natural_language_to_nrql_query and the preview trace tools, work with OAuth only, not an API key.

• New Relic MCP server: install for every agent

Axiom's hosted server lets the agent query logs and events with APL, query metrics, read the service map, and manage dashboards, monitors and notifiers. OAuth is the recommended setup. The server is hosted in the US, so query results pass through US infrastructure.

claude mcp add --transport http --scope user axiom https://mcp.axiom.co/mcp

• Read tools: queryDataset, getSavedQueries, listDatasets, getDatasetFields, queryMetrics, searchMetrics, listDashboards, checkMonitors, getMonitorHistory, services-getMap.
• Write tools: createDataset, createDashboard, updateDashboard, deleteDashboard, createMonitor, updateMonitor, deleteMonitor, createNotifier, updateNotifier, deleteNotifier. The docs list no read-only mode, so control this with the role of the user who signs in.
• Header auth for agents without browser OAuth: Authorization: Bearer PERSONAL_ACCESS_TOKEN plus x-axiom-org-id: ORGANIZATION_ID.
• Gotcha: the agent acts as your user, so Axiom says to set query cost limits on your roles.

• Axiom MCP server: install for every agent

Honeybadger's server covers errors (called faults), Insights log queries, uptime checks, check-ins, deploys and dashboards. It is the safest default here: the hosted server asks for an access level when you authorize, and the self-hosted Docker server is read-only unless you change it.

claude mcp add --transport http honeybadger "https://mcp.honeybadger.io/mcp"

{
  "mcpServers": {
    "honeybadger": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "HONEYBADGER_PERSONAL_AUTH_TOKEN", "ghcr.io/honeybadger-io/honeybadger-mcp-server"],
      "env": { "HONEYBADGER_PERSONAL_AUTH_TOKEN": "YOUR_PERSONAL_AUTH_TOKEN" }
    }
  }
}

• Main read tools: list_faults, get_fault, get_fault_counts, list_fault_notices, list_fault_affected_users, query_insights, list_deploys, list_uptime_checks, list_check_ins.
• Write tools include update_fault, create_fault_comment, create_project, delete_project, create_alarm and delete_dashboard. With read-only access they are removed from the tool list.
• Allow writes on the self-hosted server with HONEYBADGER_READ_ONLY=false. Honeybadger warns this allows deleting projects.
• EU accounts: use https://eu-mcp.honeybadger.io/mcp, or set HONEYBADGER_API_URL to https://eu-app.honeybadger.io for Docker. A US token will not work in the EU region.

• Honeybadger MCP server: install for every agent

Dynatrace runs a remote MCP server inside each environment. The agent can generate, explain and run DQL queries, investigate problems and vulnerabilities, analyze Kubernetes events, forecast timeseries, and resolve entity names. Dynatrace recommends platform tokens over OAuth, because OAuth tokens are short-lived.

https://{environment-name}.apps.dynatrace.com/platform-reserved/mcp-gateway/v0.1/servers/dynatrace-mcp/mcp

claude plugin install dynatrace@claude-plugins-official
export DT_ENVIRONMENT=https://<your-environment-id>.apps.dynatrace.com
export DT_PLATFORM_TOKEN=dt0s16.<your-token>

• The docs list capabilities, not tool names, for the remote server.
• Can it change things? The scopes the docs list for all MCP tools are read scopes plus execute scopes for AI features (such as davis-copilot:nl2dql:execute). None of them is a write scope. Access never goes beyond your own user permissions.
• Required permissions: mcp-gateway:servers:invoke and mcp-gateway:servers:read on both the user and the token. Add storage:logs:read to both if the agent should see logs.
• OAuth needs a confidential client with the Authorization Code grant. Dynamic client registration is not supported yet.
• The older local server (npx -y @dynatrace-oss/dynatrace-mcp-server) is deprecated; version 2.1.2 was its final release. It could send Slack messages, emails and events. Its README warns that Grail queries may incur costs.

• Dynatrace MCP server: install for every agent

PagerDuty's hosted server is for incident response. The agent can read incidents, past and related incidents, services, schedules, on-call, change events and log entries, and can create and update incidents, add responders and post status page updates. The self-hosted server is deprecated and its GitHub repo is archived.

PAGERDUTY_API_KEY=YOUR_API_KEY claude mcp add-json pagerduty '{"type":"http","url":"https://mcp.pagerduty.com/mcp","headers":{"Authorization":"Token token=${PAGERDUTY_API_KEY}"}}'

• OAuth: PagerDuty does not support dynamic client registration, so you create an OAuth client and pass its client ID and secret. In Claude Code, use a static callback port such as 8080.
• EU accounts: https://mcp.eu.pagerduty.com/mcp.
• Read tools include browse_incidents, browse_services, browse_schedules, browse_teams, browse_escalation_policies, browse_users, browse_change_events, browse_status_pages, browse_activity.
• Write tools: manage_incidents, manage_services, manage_schedules, manage_teams (including delete), manage_event_orchestrations, manage_alert_grouping, manage_status_pages.
• Safety: tool filtering is not available on the hosted server and the docs mention no read-only mode. PagerDuty says to limit which tools succeed with a Scoped OAuth client; otherwise access comes from the permissions of the key or user you connect with.
• Requirement: a PagerDuty account with Advanced Permissions. The docs do not say whether the Free plan qualifies.

• PagerDuty MCP server: install for every agent

• Debugging app errors: Sentry, scoped to one project. If you use Honeybadger, its self-hosted server is read-only by default and the hosted one can be authorized read-only, which is safer for an agent.
• Traces: Honeycomb (get_trace, run_bubbleup, get_service_map). New Relic and Grafana Tempo are options if your traces already live there.
• Logs: Axiom for a large free tier and APL queries; Grafana (Loki) or New Relic if your logs are already there.
• Grafana stacks: Grafana Cloud MCP if you are on Grafana Cloud; the open source mcp-grafana with --disable-write for self-hosted Grafana.
• Incident response: PagerDuty for paging, on-call and status pages. Grafana covers incidents and OnCall if you use Grafana IRM.
• Already on New Relic or Dynatrace: use their hosted servers; both focus on queries and analysis.
• For any server that can write, connect with a least-privilege user or token, and keep tokens in environment variables, not committed files.
• All commands, tool names and prices were checked against vendor docs and pricing pages on 2026-10-23.

• Set up MCP servers, rules and permissions in one paste
• Add an MCP server in every AI coding agent