GenZAgents for Cursor — agent receipts in the AI-first IDE

A Cursor agent run touches dozens of files in seconds. The accountability gap between "what did the AI just change" and "why" is the largest in any AI-tool surface. GenZAgents closes that gap with per-run receipts: every Cursor Agent invocation produces a receipt listing the files touched, the diff hash, the model, the runtime cost, and the engineer who pressed go.

Cursor has three distinct surfaces that can mutate code: Composer (multi-file edits with planning), Agent mode (longer autonomous loops), and inline Cmd-K (single-file inline edits). The GenZAgents MCP server is invoked from any of them — if Cursor calls our tools, we emit a receipt. The exception is pure inline tab-complete (no agentic loop), which today doesn't produce a receipt; v0.8 will add a Cursor-specific completion-event hook for full coverage.

Senior engineers often hop between Cursor (for fast iteration) and Claude Code (for longer sessions with deep file-system tool calls). The receipts unify them: every session emits a receipt tagged with the same agent DID and project regardless of which IDE ran it. The /agents/[did]/restore endpoint then lets you reconstruct the work history as one feed across both IDEs.

Cursor Agent burns through tokens. Most teams have no per-engineer accountability for the spend — the invoice is one line for the whole team. GenZAgents tags every receipt with the cost in USD + the model variant used + the engineer DID. The /settings/billing dashboard rolls these up by team / project / engineer. Knowing that 3 engineers drove 70% of last month's Cursor spend is the first step to optimising it.

EU AI Act §50 and the September 2026 SOC 2 Trust Services Criteria addendum both require "evidence of the AI-assisted nature of work products." A code commit that the AI wrote needs a paper trail. The GenZAgents per-commit receipt + the per-Cursor-Agent-run receipt together provide that trail: "this commit was produced by a Cursor Agent run with claude-3-5-sonnet, runtime cost $0.34, supervised by engineer did:genz:abc…".

For one developer: paste the config block into Cursor Settings. For a team: push @genzagentsio/setup via your MDM, with the org install token in /etc/genzagents.env. The setup CLI auto-detects Cursor among the installed editors, writes the MCP block, and never overwrites existing user settings. Engineers see GenZAgents working without ever opening the config dialog.

Cursor → Settings → Features → MCP → Add new MCP server. Paste:

{
  "command": "npx",
  "args": ["-y", "@genzagentsio/mcp-server"]
}

Give it a name (genzagents) and save.

Cursor Agent runs, Composer runs, Cmd-K sessions where MCP tools were called.

Open Cursor → Cmd-K → "use the genzagents tool to draft a receipt" → confirm draft on /dashboard.