What 1,137 agent writes taught this MCP server author about tool scoping and safety
QuanTradin · reddit · 2026-09-04
The author of a production MCP server managing a paper trading desk (22 tools) shared audit-ledger lessons: since Aug 19 agents made 1,137 writes, deployed 781 bots, and lost $1.34M in paper money — 773 of the 781 never called the backtest tool first, proving agents skip optional tools.
Controls that held:
- Key-pair auth bound to one desk; pointing it elsewhere is refused with no cross-desk tool to negotiate.
- Scope lives on the key, not in the prompt (read/backtest/deploy); out-of-scope calls get a 403 with a message framed as a user-set limit, not a judgment.
- Asymmetric kill switch: a tool reads freeze state, but unfreezing is a session-only route — only a human in a browser can re-enable agents.
Two mistakes admitted: the ledger only records successful writes (403s raise before logging), so there's no record of attempts; and zero of 33 minted keys used the safe backtest-only default. Works via Claude Code and Claude Desktop (mcp-remote); the claude.ai web and API MCP connectors are OAuth-only and unsupported.
More from coding & agent
- GPT-6 Astra builds explorable 3D city scenes in Unity, 1.9x faster than GPT-5.6 Sol — ChrisGPT · 2026-09-04
- Dev builds a working macOS 27 simulator with Astra in 75 minutes, cloud sync included — pvncher · 2026-09-04
- Opinion: AI Can Amplify Technical Debt — craigmullins · 2026-09-04
- Sandbox readiness is the new bottleneck in agentic apps; git as state primitive — w_hgm · 2026-09-04
- OSWorld Creator: From Operator to GPT-6 Astra on OSWorld 2.0 — TianbaoX · 2026-09-04
- Theo: GPT-6 Astra one-shots a 3D browser game, world-class at Blender and spatial reasoning — charliermarsh · 2026-09-04