SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI codingbyok-observability-mcp

byok-observability-mcp

Observability across Grafana, Prometheus, Kafka UI and Datadog using your own keys.

Open website

About this tool

Purpose

byok-observability-mcp is used to query the local observability stack from Claude Code, Codex, or any MCP client, and the data never leaves the local machine. It is aimed at developers and operations personnel who need to connect the query capabilities of monitoring and logging platforms to AI clients.

Main capabilities

Supported backends include:

  • Grafana: enabled when GRAFANA_URL and GRAFANA_TOKEN are set.
  • Prometheus: enabled when PROMETHEUS_URL is set.
  • Kafka UI: enabled when KAFKA_UI_URL is set.
  • Datadog: enabled when both DD_API_KEY and DD_APP_KEY are set; it actually proxies the official Datadog MCP server. The default toolset is core, apm, alerting; setting DD_TOOLSETS=all loads all of them.

In addition, a tool is always provided that can be used to check the connectivity of all configured backends. All tools are read-only and do not perform write operations on any backend.

Partial configuration on demand is supported: configure only the backends you use, and tools for unconfigured backends will not be exposed.

Scheduled reports are supported: automated observability summaries can be sent to Slack on a schedule without running a Claude or Codex instance. The command collects data, classifies it by severity, formats and sends a Slack message, and then exits, completely stateless.

How to use

Method A — Interactive wizard (recommended): Run the wizard once, answer a few questions, and a ready-made .mcp.json is generated. The wizard will: let you choose the backends to configure; ask for credentials by service; test connectivity with real endpoints before writing; write .mcp.json to the project root or ~/.claude/ (your choice). Then start Claude Code. No cloning, building, or env file is needed, and it can be completed within 60 seconds.

Method B — Manually configure .mcp.json: Create .mcp.json in the project root and include only the required backends. If credentials would enter git, you can use the ${VAR} approach (Configuration → Method B). After starting Claude Code, it will automatically read .mcp.json, with no need for claude mcp add and no build step.

Supported clients: Both Claude Code and Codex automatically read .mcp.json, and it can also be added via CLI.

Obtaining credentials:

  • Grafana: Administration → Users and access → Service accounts → Add service account (set Role to Viewer) → Create → Add service account token → Generate token, and copy the token starting with glsa_.
  • Prometheus: no credentials are needed when there is no authentication; set them accordingly when using basic auth.
  • Kafka UI: no credentials are needed when there is no authentication; provide an account when login is required.
  • Datadog: the API key is at Organization Settings → API Keys → New Key; the Application key is at Organization Settings → Application Keys → New Key; DD_SITE must match your Datadog login URL.

Configuration methods:

  • Method A: Fill in credential values directly in .mcp.json (simplest); when sharing the repository, add .mcp.json to .gitignore.
  • Method B: Use ${VAR} placeholders in .mcp.json, put the real values in .env (add it to .gitignore), and load that env when starting Claude. Ready-made helper scripts are included, as well as a template .mcp.json.example containing all variables. It can also be added to ~/.claude.json.

Scheduled report configuration:

  1. Obtain a Slack Incoming Webhook URL: api.slack.com/apps → Create New App → From scratch → turn on Incoming Webhooks → Add New Webhook to Workspace → select a channel → Allow → Copy the Webhook URL.
  2. Set the environment variables.
  3. Run a one-time report validation; you should see a message in the Slack channel within seconds.
  4. Schedule with cron: edit crontab and add the corresponding line; you can put all environment variables in .env and source them in the cron command. On macOS you can also use launchd (runs at login and remains effective after restart); create ~/Library/LaunchAgents/com.observability-mcp.report.plist. To stop, use launchctl unload.

Example prompts: Supports single-backend queries, cross-backend queries, and the incident response scenarios in v0.2.0+.

Security

  • All tools are read-only and perform no write operations on any backend.
  • Credentials are read only from environment variables and are never logged or sent to Anthropic. Tokens in all error messages are redacted.
  • TLS certificate verification is enabled by default.
  • The MCP process runs locally, and these URLs enter Claude's context window only when you enter your infrastructure URLs into the chat.
  • The documentation provides least-privilege recommendations.

Applicable scenarios

Suitable for development and operations scenarios where you want to perform read-only queries against local observability platforms (Grafana, Prometheus, Kafka UI, Datadog) through MCP clients such as Claude Code and Codex, and want credentials and data to remain local; also suitable for teams that need to push observability summaries to Slack on a schedule.

Limitations

The material does not provide specific platform compatibility scope, pricing and billing information, performance metrics, or effect guarantees; some backends are enabled only when the corresponding environment variables are set; Datadog capabilities depend on the official Datadog MCP server. Specific tested version numbers, the complete environment variable list, and example command details not given in the material are not provided for now.

Related tools

Browse category