cursor-delegate

{
  "name": "cursor-delegate",
  "description": "Delegate self-contained coding/research tasks to the Cursor CLI (cursor-agent) from inside Claude Code, so they run on the Cursor subscription's quota instead of Claude's. One task via the cursor-delegate skill, or a whole fleet via cursor-orchestrate (Claude plans and reviews, Cursor executes in parallel). Workers run fully trusted (no approval prompts, like Claude Code subagents). Long tasks run legged: ~4-minute legs resumed on one session, so flaky networks/VPNs cannot kill the stream. Auth is key-based and deterministic (agent-keys.json default entry); runs self-terminate the moment their result lands; every session is resumable so no work is ever lost. Workers can author their own report cards (*-card.html) that render as chat widgets via the readable plugin at near-zero context cost. Model routing follows Cursor's two quota pools: first-party models (auto, composer-*, cursor-*) for mechanical work, API pass-through models (claude-*, gpt-*) kept for prose and judgment, with a pre-flight probe before big fan-outs. Agent calling, not a model-backend swap. Also carries copy-writing-fa: Persian web copy (articles, product descriptions, landing pages, microcopy, SEO meta) as a five-stage pipeline, with fa-lint, a dependency-free Persian linter that enforces ZWNJ, Persian punctuation and digits, administrative words, ad clichés and voice as a real exit code, and measures whether the copy reads human (sentence rhythm, repeated openers, repeated phrases) against thresholds calibrated on a real human corpus.",
  "version": "1.9.1",
  "author": {
    "name": "Seyed M. Keramati",
    "url": "https://github.com/SMKeramati"
  },
  "mcpServers": {
    "cursor-delegate": {
      "command": "node",
      "args": [
        "${CLAUDE_PLUGIN_ROOT}/mcp/cursor-mcp.js"
      ],
      "env": {
        "CURSOR_RUN_BIN": "${CLAUDE_PLUGIN_ROOT}/scripts/cursor-run.sh"
      }
    }
  }
}