mcp server

pngr mcp turns the CLI into a Model Context Protocol server. AI agents discover the available tools and their parameter schemas automatically — so an agent can manage your monitoring through natural language with no prompt engineering or doc-reading.

setup

Install the CLI, then point your agent client at it. For Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "pngr": {
      "command": "pngr",
      "args": ["mcp"],
      "env": { "PNGR_TOKEN": "pngr_pat_…" }
    }
  }
}

Replace the placeholder with a personal access token from /settings/api-tokens, then restart the client. Cursor, Windsurf, Zed, and other MCP-compatible clients take an equivalent config in their own settings. The /settings/mcp page in the web UI has a copy-paste version of this snippet and the live list of exposed tools.

transports

• stdio (default) — for local agents that spawn pngr mcp as a subprocess. This is what the config above uses.
• SSE (pngr mcp --transport sse) — for remote or hosted agents that need a persistent HTTP endpoint.

the tools

Each operation maps to one tool, named noun_verb for discoverability — monitor_list, monitor_create, monitor_check, incident_list, channel_test, rule_list, dashboard_summary, and so on. Tools are tagged by scope (read, write, admin). The /settings/mcp page shows the exact set your version exposes — it's the authoritative list, since the tool set grows over time.

The stack-as-code tools deserve a special mention:

• stack_dump exports the whole org as one YAML artifact, so an agent edits a single file instead of issuing dozens of calls.
• stack_apply returns the diff first; the agent relays the changeset, then re-calls with confirm: true to execute — the same preview-then-confirm safety the CLI and /stack page enforce.