What it is
The Grafana MCP Server connects agents to a Grafana instance and its ecosystem. Its README lists 129 tools: searching, summarising and patching dashboards; querying Prometheus, Loki, Tempo, Pyroscope and several SQL and cloud data sources; managing alert rules, silences and contact points; and working with Grafana Incident and OnCall. Forty of them, including the admin, panel-query, SQL and Grafana Assistant tools, are off until enabled.
Install with uvx mcp-grafana and set GRAFANA_URL plus a service account token. It works with self-hosted Grafana 9.0 or later and with Grafana Cloud, runs as a binary, in Docker or via Helm, over stdio, SSE or Streamable HTTP, and reached version 2.1.1 on 20 October 2026.
Grafana Cloud customers also have a hosted option at https://mcp.grafana.com/mcp, which uses OAuth 2.1 and acts with the signed-in user's permissions. Each person who connects that way counts as an active Grafana Assistant user for billing.
Install
uvx mcp-grafanaOur review
For on-call engineers, this server lets an agent go from an alert to the relevant dashboard, PromQL query, log lines and trace in one conversation, then check who is on call and add a note to the incident. With 129 tools across the stack, its coverage of Grafana's own products is close to complete.
Context management is why it works in practice. get_dashboard_summary, and get_dashboard_property with a JSONPath expression, return only the parts of a dashboard the agent needs, and the patch tool edits without resending the full JSON; the README flags where a tool still returns large payloads. Forty tools are disabled by default, so a fresh install offers roughly 89 rather than everything.
The permission controls are finer than most. --disable-write removes write tools, and it also removes query_sql and query_influxdb, because those run whatever they are given and could drop a table if the data source credentials allow. --disable-query keeps only discovery tools, and --loki-enforced-matchers adds fixed label filters to every Loki query, a workaround for open-source Loki's lack of per-user label controls. Two defaults deserve attention: the Docker image starts in SSE mode on 0.0.0.0:8000 and only logs an error if no caller token is set, and anonymous usage statistics go to Grafana Labs every four hours unless disabled.
The hosted Grafana Cloud server changes the buying question rather than the engineering one. It removes token setup and scopes access to each user's RBAC, but every person who connects counts as an active AI user, even if they never touch Assistant's AI features. Three come with the $19 Pro platform fee and each extra one starts at $20 a month, so a 20-engineer on-call rotation adds about $340 a month that the self-run server avoids.
Where it earns its place
An on-call SRE at a 50-engineer company on self-hosted Grafana
Work out why checkout latency spiked at 02:00
The agent finds the service dashboard, runs query_prometheus and query_loki_logs over the window, pulls the matching Tempo trace and drafts an incident note. A Viewer service account with --disable-write keeps it read-only, and the server costs nothing.
An engineering manager on Grafana Cloud Pro
Give 12 engineers agent access to observability data without handing out tokens
Point Claude Code or Cursor at https://mcp.grafana.com/mcp; each engineer authorises in the browser and keeps their own RBAC. The nine engineers beyond the three included cost from $180 a month as active AI users.
A platform engineer cleaning up a noisy alerting setup
Find flapping alert rules and silence or retune them in bulk
alerting_rules_read lists rules and their states, alerting_silences_write creates silences and alerting_rules_write updates thresholds. Enable write tools only for this job, with a service account scoped to alerting.
Before you commit
- Setup
- Create a Grafana service account token, then add uvx mcp-grafana with GRAFANA_URL and GRAFANA_SERVICE_ACCOUNT_TOKEN to Claude Desktop, Cursor or VS Code; minutes to the first dashboard search. Grafana Cloud users can instead run claude mcp add grafana --transport http https://mcp.grafana.com/mcp and sign in.
- Integrations
- 129 tools, about 89 on by default: dashboards, folders, data sources, Prometheus, Loki, Tempo, Pyroscope, alerting, Incident, OnCall, annotations, snapshots, panel rendering, docs search and a raw grafana_api_request tool. Optional groups add CloudWatch, Elasticsearch, InfluxDB, SQL and Agent Observability.
- Data and privacy
- Access equals the service account's Grafana role, or the user's RBAC on the hosted server. grafana_api_request can call any Grafana HTTP API the token permits and bypasses Loki label enforcement unless --disable-api is set. Log lines returned to the model are untrusted text.
- Cost at scale
- The self-run server is free; its queries add load to Prometheus and Loki and, on Grafana Cloud, to usage-based bills. Hosted MCP users count as active AI users: 3 included with Pro's $19 platform fee, then from $20 each a month with 40M tokens per user.
- Limits
- Grafana 9.0 or later is required. get_dashboard_by_uid still floods context on large dashboards. The hosted server works only with Grafana Cloud, needs Assistant terms accepted, and supports Streamable HTTP but not SSE.
What's good
- 129 tools across dashboards, metrics, logs, traces, alerting and on-call
- Summary and JSONPath tools keep dashboards out of context
- Read-only and query-free modes, plus Loki label enforcement
- Hosted OAuth option for Grafana Cloud
What's not
- Docker image listens on 0.0.0.0 without a caller token by default
- Hosted server bills each connected user from $20 a month
- Incident, OnCall and Assistant tools need their plugins
- Usage statistics on by default
Best for
SRE and platform teams on Grafana or Grafana Cloud who want agent help during investigations.
Alternatives
Grafana MCP Server: common questions
- Is Grafana MCP Server worth it?
- Grafana has thought harder than most vendors about context size and permissions, and the result is a server an on-call engineer can leave attached through an incident. We rate it 4.3 out of 5.
- How much does Grafana MCP Server cost?
- Self-run server free (Apache-2.0). Hosted Grafana Cloud MCP counts each user as an Assistant active AI user: 3 included with Pro's $19 platform fee, then from $20 per user a month.
- What are the downsides of Grafana MCP Server?
- Docker image listens on 0.0.0.0 without a caller token by default; hosted server bills each connected user from $20 a month; incident, OnCall and Assistant tools need their plugins; usage statistics on by default.
- Who is Grafana MCP Server best for?
- SRE and platform teams on Grafana or Grafana Cloud who want agent help during investigations.
- How much does Grafana MCP Server cost for a team or in production?
- The self-run server is free; its queries add load to Prometheus and Loki and, on Grafana Cloud, to usage-based bills. Hosted MCP users count as active AI users: 3 included with Pro's $19 platform fee, then from $20 each a month with 40M tokens per user.
- Is Grafana MCP Server safe for company data?
- Access equals the service account's Grafana role, or the user's RBAC on the hosted server. grafana_api_request can call any Grafana HTTP API the token permits and bypasses Loki label enforcement unless --disable-api is set. Log lines returned to the model are untrusted text.
- What are the best Grafana MCP Server alternatives?
- The closest alternatives we have reviewed are Sentry MCP and ClickHouse MCP Server.
Answers checked Oct 22, 2026.
Make Grafana MCP Server? Show this rating on your site
