robotyx-mcp

Stdio MCP server that wraps Robotyx's REST API. Lets Claude Code / Claude Desktop browse recorded workflows, refine their TestSteps into grouped functional steps ("clicks 1-7 → Log in as admin"), and push the result back — without leaving the chat.

No new business logic lives here. Every tool is a thin wrapper around an existing endpoint the Robotyx web UI also uses.

Tools

| Tool | Wraps | Notes | |------|-------|-------| | list_workflows | GET /api/workflows | Summary rows. | | get_workflow | GET /api/workflows/:id | Screenshots stripped by default. Pass include_screenshots: "all" or "indices" + screenshot_step_indices: [1,3,5] when you need visual context. | | list_testsuites | GET /api/testsuites | | | get_testsuite | GET /api/testsuites/:id | Includes cases + their steps. | | create_testsuite | POST /api/testsuites | | | add_test_case | POST /api/testsuites/:id/cases | Optionally links a Workflow id. | | set_test_steps | POST /api/testsuites/:id/cases/:caseId/import-mapping | The refine path. Body shape mirrors the UI's Excel import — pass workflow_step_ids (full ordered list from get_workflow) + mappings: [{description, expected_result, workflow_step_indices: [1,2,3]}]. | | merge_step | POST .../steps/:stepId/merge | | | split_step | POST .../steps/:stepId/split | | | create_defect | POST /api/defects | |

Wire-up — Claude Code

Drop a .mcp.json in your workspace root (already done for ai-automation-platform/ and footprint-platform/):

{
  "mcpServers": {
    "robotyx": {
      "command": "npx",
      "args": ["-y", "tsx", "C:\\Users\\…\\ai-automation-platform\\robotyx-mcp\\src\\index.ts"],
      "env": {
        "ROBOTYX_URL": "http://localhost:3001",
        "ROBOTYX_API_KEY": "${ROBOTYX_API_KEY}"
      }
    }
  }
}

Then in PowerShell (once per machine):

[Environment]::SetEnvironmentVariable("ROBOTYX_API_KEY", "rx_…", "User")

Restart Claude Code so it picks up the new MCP. Verify with /mcp — you should see robotyx: ✓ connected (10 tools).

Wire-up — Claude Desktop

Add the same mcpServers block to %APPDATA%\Claude\claude_desktop_config.json and restart the app.

Typical refine loop

1. list_workflows → pick a workflow id.
2. get_workflow (no screenshots) → AI proposes a grouping based on step descriptions + action types.
3. get_workflow with include_screenshots: "indices" and the ambiguous step indices → AI disambiguates visually.
4. Either add_test_case (linking the workflow) or pick an existing one from get_testsuite.
5. set_test_steps with the proposed mapping → done.

Token budget

A 30-step browser recording with full-page PNGs is ~200-400k tokens with include_screenshots: "all". Always start with "none", then use "indices" to fetch only the screenshots you need. The server strips footprint_* bookkeeping keys (raw a11y dumps, console logs) unconditionally.