Brokerage MCP Postmortem: 4 Pitfalls Making Trader Leaderboards Safe for Agents
Accomplished_Fun_408 · reddit · 2026-09-24
The RockFlow team (HK/US stocks & options brokerage) shares 4 lessons from exposing a read-only trader leaderboard and public-holdings MCP to LLM agents:
- Ratios, never amounts: positions return only weights and return percentages; agents kept trying to "estimate" hidden amounts, so tool descriptions now forbid it
- IDs as strings: 10–20 digit trader IDs silently lost trailing digits when models passed them as JSON numbers, looking up the wrong person
- Two units in one payload: score arrives as a percentage (22.5 = +22.5%) while yearlyYield/maxDrawDown are decimals (0.35 = +35%); agents compared them directly until descriptions spelled this out
- Nicknames are prompt-injection surface: free-text trader nicknames are now explicitly treated as names only
The MCP is live for Claude/Cursor/Codex (paper account, OAuth only), and the author asks how others handle confirmation for irreversible tools.
More from coding & agent
- Treating AI video as a production pipeline: building a reusable cinematic producer agent in CREAO — FellMentKE · 2026-09-24
- Free MCP connector serves UN SDG indicators for all 17 goals by country and year — modelcontextprotocol · 2026-09-24
- Gemini MCP server brings search, deep research and image/video generation into Claude — modelcontextprotocol · 2026-09-24
- Chose rustpython-parser over tree-sitter after measuring both on real LLM output — SprayPuzzleheaded533 · 2026-09-24
- More compute made one agent 13x faster, another just 4%: agents' bottlenecks are task-dependent — alex_verem · 2026-09-24
- At least 4 WebKit WebView 'vibe-coded browsers' appeared in 24 hours — what triggered the wave? — uwukko · 2026-09-24