@nijam/mcp-server

MCP server for Nijam — lets any MCP-capable agent (Claude Code, Claude Desktop, Cursor, …) query your Playwright test runs: what failed, why, since when, and what's flaky.

Ask things like "why is the checkout suite red?", "is cart updates quantity flaky?", or "when did this test start failing?" — the model answers from your real run history.

Setup

You need a Nijam read key (nij_rk_…): create one in your dashboard under Organization settings → Secret keys, choosing the Read (MCP) type. Read keys are read-only by construction — they can never upload or change anything — and always cover the whole organization, so the agent can resolve any project by name or slug. (Ingestion keys — nij_sk_…, the ones in your CI — are write-only and are rejected by the read API, so a leaked CI key exposes no data.)

Claude Code

claude mcp add nijam -e NIJAM_API_KEY=nij_rk_... -- npx -y @nijam/mcp-server

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "nijam": {
      "command": "npx",
      "args": ["-y", "@nijam/mcp-server"],
      "env": { "NIJAM_API_KEY": "nij_rk_..." }
    }
  }
}

| Env var | Required | Default | | | --- | --- | --- | --- | | NIJAM_API_KEY | yes | — | Read key (nij_rk_…) from the dashboard. | | NIJAM_API_URL | no | https://api.nijam.dev | Self-hosted / local API base URL. |

Tools

get_projects is the entry point: it lists the projects the key can access as { id, slug, name }. Every other tool's project parameter accepts the id, the slug, or the name — so "check the Web Checkout project" resolves without you ever pasting a UUID.

| Tool | What it answers | | --- | --- | | get_projects | What projects can I see? (call first — resolves "my project" into an id/slug) | | get_latest_run | What's the state of the latest run (optionally on a branch)? | | list_failing_tests | Which tests failed in this run (or the latest run), and with what error? | | get_failure | Full detail of one failure: every attempt, error, screenshots/video/trace links. | | is_test_flaky | Verdict + evidence: how often did this test flake recently? | | get_run | One run's aggregate stats and per-spec-file breakdown. | | list_runs | Run history, filterable by status/branch, paginated. | | get_test_history | One test across the last 30 runs — spot when it started failing. | | list_flaky_tests | The flakiest tests of the project, ranked by flake count. |

Reads only — read keys cannot write, and the server never mutates anything.

Notes

• Slugs are derived from the project name ("Web Checkout" → web-checkout). If two projects share a slug, the tools say so and ask for the id.
• Trace artifacts are linked via short-lived (15 min) presigned URLs, minted on demand.
• Errors come back as tool errors with the API's own message, so the model can self-correct (e.g. unknown slug → the list of valid slugs).

License

MIT