@belvador/mcp

The Belvador MCP server over stdio, for AI agents that run on your machine (Claude Code, Cursor, VS Code, Codex, Windsurf…). It offers the same tools as the remote server at https://api.belvador.dev/mcp — read-only by default — and reads node data (issues, events, traces, logs) straight from edge.belvador.dev with short-lived tickets, so that data never passes the Belvador API.

{
  "mcpServers": {
    "belvador": {
      "command": "npx",
      "args": ["-y", "@belvador/mcp"],
      "env": { "BELVADOR_TOKEN": "blv_pat_…" }
    }
  }
}

• Token: a read-only personal or service token (blv_pat_ / blv_svc_) from Settings → API tokens, in BELVADOR_TOKEN, or in the OS keychain (service belvador.dev, account mcp):
  • macOS: security add-generic-password -s belvador.dev -a mcp -w
  • Linux: secret-tool store --label=Belvador service belvador.dev account mcp
• Toolsets: --toolsets=investigate,reliability,docs (the default); the token's scopes narrow them further. Add configure to change settings (search_operations, describe_operation, plan_change, apply_change, get_operation); it needs a token with write scopes. Sensitive changes — keys, members, tokens, alert destinations, node changes — answer approval_required with a belvador.dev link: a person approves with their authenticator code, and new secrets are shown only to them, never to the agent.
• Untrusted data: everything that came from your apps is returned under untrusted, cleaned and size-limited; the server's instructions tell the agent not to follow text inside it. No tool opens URLs found in data.
• Logs: diagnostics go to stderr and never contain the token or tool results.

Prefer no local process? Use the remote server https://api.belvador.dev/mcp with OAuth or a token. Docs: https://belvador.dev/docs/mcp · Node ≥ 20.10 · MIT licensed.