grafana-mcp-adapter

Read-only MCP server that exposes Grafana dashboards, datasources and metric/log queries as tools, for use by skills in the TICKETS workspace.

This is a private, standalone adapter: its own git repo, its own .env. It is not part of ia-tooling and does not depend on its stack — it only makes HTTPS calls out to the Grafana HTTP API (${GRAFANA_BASE_URL}/api). It was scaffolded by cloning the structure of fathom-mcp-adapter.

Architecture

The adapter is the MCP server. Internally it calls the Grafana HTTP API directly (just as the Fathom adapter calls Fathom's REST API). MCP servers are not chained to one another.

• Auth: a Grafana service account token sent as a Bearer token in the Authorization header. Scope it to a read-only role (Viewer). Create the service account + token in Grafana → Administration → Service accounts.

• Most tools return raw payloads (dashboard JSON, datasource lists). The exceptions are the high-volume ones: grafana_query returns a compact per-series digest by default (a full up query is ~8 MB of frames, far past what an MCP context wants), and grafana_logs_link returns a capped preview. Both can be pushed back toward raw (raw=true / a higher limit).

• Read-only by design. Note that grafana_query is a POST (Grafana's /api/ds/query is POST-shaped) but only reads metrics/logs.

Related MCP server: poly-observability-mcp

Tools

grafana_query response shape

The raw /api/ds/query response carries a full timestamp+value array per series, and a query like up can return thousands of series (~8 MB). By default the tool collapses each series to its labels + a numeric digest and caps the list:

{
  "results": {
    "A": {
      "status": 200,
      "series_count": 3085,        // total series before capping
      "series": [                  // capped to maxSeries (50)
        { "labels": { "job": "..." }, "count": 60, "first": 1, "last": 1, "min": 0, "max": 1, "avg": 0.98 }
      ],
      "truncated": 3035            // how many series were dropped from `series`
    }
  }
}

Pass raw=true to get the full (potentially very large) frames instead.

grafana_logs_link

Identify a customer/component with free text (client='april', component='gateway'); it matches case-insensitively against the service_name label, which on this instance encodes both (e.g. graviteeio-ae-april-rec-engine). Returns { query, explore_url, range, preview_count, preview } — paste explore_url (a permanent Grafana 11+ Explore link) into the ticket. The default range is the last hour; widen with from/to. When nothing matches, it returns close service_name values as suggestions so typos like aprl → april surface.

Examples (how a user asks for it)

Just ask in plain language — the agent maps it to the client / component / from / to / line_filter arguments for you.

"Give me the last hour of API gateway logs for Northwind." → { "client": "northwind", "component": "gateway" }

"Show me the engine logs for Contoso over the last 6 hours." → { "client": "contoso", "component": "engine", "from": "now-6h" }

"Find the gateway errors for Globex in the last 3 hours, give me up to 200 lines." → { "client": "globex", "component": "gateway", "line_filter": "error", "from": "now-3h", "limit": 200 }

"I need the UI logs for Initech during yesterday's incident between 10:00 and 11:00." → { "client": "initech", "component": "ui", "from": "<epoch ms 10:00>", "to": "<epoch ms 11:00>" }

"Give me the production gateway logs for Northwind." → { "client": "northwind prod", "component": "gateway" }

(The environment — prod, rec, dev — isn't a separate argument: it lives inside service_name, so just fold it into client as another word. Both client and component are matched as case-insensitive substrings with .* between them, so northwind prod matches …-northwind-prod-….)

Each call returns explore_url — paste it into the ticket as a permanent link — plus a capped preview of the newest matching lines.

Setup

npm install
cp .env.example .env   # then fill in GRAFANA_BASE_URL and GRAFANA_TOKEN
npm test
npm start              # speaks MCP over stdio

GRAFANA_TOKEN must come from .env (which is git-ignored) — never hardcode it.

Wiring into the TICKETS MCP session

Build the image and add a grafana entry to the TICKETS .mcp.json, alongside fathom / zendesk / vectordb / github. It only needs HTTPS egress.

docker build -t grafana-mcp-adapter .

// .mcp.json (TICKETS)
{
  "mcpServers": {
    "grafana": {
      "command": "docker",
      "args": ["run", "--rm", "-i", "--env-file", "/ABSOLUTE/PATH/TO/.env", "grafana-mcp-adapter"]
    }
  }
}