Claude stays your driver. When a task suits Cursor better — a second opinion, parallel work, a long unattended fix — Claude delegates it to cursor-agent, tracks it as a background job, and reports the result back.

You ──▶ Claude Code ──/cursor:delegate──▶ cursor-agent (background)
                  ◀── result + session id ──┘

• 🚀 Background handoff — fire a task, keep working; Claude polls and reports.
• 📊 Live progress — watch a running job's activity, tool calls, changed files, tokens, and cost as it works.
• 🔔 Completion-aware — a Stop hook tells Claude (and pings your desktop) the moment a background job finishes — no polling.
• 🔍 Reviewable — /cursor:diff <id> shows the git diff of exactly what a job changed.
• 🔁 Resumable + self-healing — continue a session with new instructions, or --retry N to auto-resume on crash.
• 📈 Statusline — optional one-line ambient summary of active jobs.
• 🤖 Self-rescue — a cursor-rescue subagent Claude can invoke when stuck.
• 🧪 Tested — 84 mocked tests + opt-in live smoke test, CI on every push.

Contents

• Prerequisites
• Install
• Quickstart
• Commands
• How it works
• Risks — read this
• Configuration
• Tests
• License

Prerequisites

Authenticate Cursor once: cursor-agent login (or export CURSOR_API_KEY).

Install

From inside Claude Code, add the marketplace and install the plugin:

/plugin marketplace add akash-viven/cursor-plugin-cc
/plugin install cursor@akash-viven

Verify everything is wired up:

/cursor:setup

Updating — pull the latest version anytime:

/plugin marketplace update akash-viven

[!TIP] If commands don't appear after install or update, refresh the runtime with /reload-plugins. Still stuck? Remove and re-add: /plugin marketplace remove akash-viven then /plugin marketplace add akash-viven/cursor-plugin-cc.

Quickstart

# 1. Hand a task to Cursor (returns a job id, runs in the background)
/cursor:delegate refactor src/auth.ts to use async/await and add tests

# 2. Check progress — bare for the list, or pass an id for a live card
/cursor:status
/cursor:status <id>      # activity timeline, tool calls, changed files, elapsed

# 3. Read the result once it's done (shows tokens + cost)
/cursor:result <id>

# 4. Review exactly what changed
/cursor:diff <id>

# 5. Iterate on the same Cursor session
/cursor:resume <id> now wire it into the login route

When a background job finishes, a Stop hook surfaces it to Claude automatically and fires a desktop notification — you don't have to keep checking status.

Job ids accept a unique prefix — /cursor:result 1718 works if it's unambiguous.

Commands

Flags on delegate / resume: --model <m> (override model), --retry <n> (auto-resume on crash/failure up to n times), --worktree (isolate edits in a throwaway git worktree, delegate only).

A cursor-rescue subagent ships alongside, so Claude can self-delegate when it gets stuck.

Completion notifications

A bundled Stop/SubagentStop hook scans for background jobs that finished since the last turn and feeds a one-line summary back to Claude as context — so Claude reacts to completions on its own. It also fires a best-effort desktop notification (macOS osascript / Linux notify-send). Tune or disable via env (see Configuration).

Statusline (optional)

Show active jobs in your Claude Code statusline — e.g. ◐ 2 cursor · 3m 12s, empty when idle. Add to your settings.json, pointing at the installed plugin path:

"statusLine": {
  "type": "command",
  "command": "~/.claude/plugins/marketplaces/akash-viven/plugins/cursor/scripts/statusline.sh"
}

How it works

cursor-agent has no long-lived daemon, so each handoff is a detached subprocess. delegate.sh runs:

cursor-agent -p "<task>" --output-format stream-json \
  --force --trust --sandbox disabled --workspace "$PWD"

…detached, streaming NDJSON events to output.json as it works. /cursor:status <id> parses that live stream on demand — no daemon, no polling loop — to show the model, elapsed time, recent activity, tool calls, changed files, and running token/cost usage. When the job exits, the wrapper reads the final result event (.result, .session_id, .is_error, .duration_ms, .usage) into per-job files. With --retry N, a failed or crashed run is re-attempted — resuming the captured session so the agent continues rather than restarting cold.

Job state lives in ~/.cursor-plugin/jobs/<id>/:

cwd  prompt  model  started  status  pid  output.json  exit_code  err.log
result.txt  session_id  duration_ms  tokens_in  tokens_out  cost_usd
retry_max  attempts  announced

status is one of:

⚠️ Risks — read this

[!WARNING] Delegated jobs run unattended with write access. Treat this like handing your repo to an autonomous agent.

• Unattended file edits. Jobs run with --force --trust --sandbox disabled. cursor-agent can write files and run shell commands in your repo without prompting. Only delegate in repos you trust, and keep a clean git tree so you can diff/revert.
• Concurrent edits. Multiple background jobs (or a job + you) editing the same files can clobber each other. Use --worktree to isolate a risky job in ~/.cursor/worktrees/.
• Quota / loops. Two AI agents can burn usage fast. Don't wire this into automatic loops.

Configuration

Environment variables (all optional):

Tests

Pure-bash harness with a mocked cursor-agent — no quota used:

tests/run.sh                 # 84 unit + integration tests
CURSOR_LIVE=1 tests/run.sh   # also runs one real cursor-agent smoke test

CI runs manifest validation, shell syntax checks, and the mocked suite on every push and PR.

License

Apache-2.0 © Akash Solanki. Not affiliated with Anysphere (Cursor) or OpenAI.