Managing Executable Tools in MCP Production Environments
Future_AGI · reddit · 2026-07-14
As a follow-up to a previous post, the author shares lessons learned from migrating an MCP server from stdio to a hosted environment, specifically focusing on running tools that execute real actions in production.
Key takeaways include:
- Separate governance for read and write tools: While read-only tools are easy to integrate, state-changing tools (running evals, adding guardrails, generating synthetic sets, writing to datasets) can cause issues if they share overly broad permissions with read tools. The solution is to implement key-based allow/deny lists.
- Don't just intercept request parameters: Beyond checking tool call inputs, it's more effective to inspect the output after execution to ensure it is grounded, policy-compliant, and structurally sound. Even valid calls returning garbage results must be blocked.
- Tools should guide the agent's next step: Returning a "score + next step suggestion" propels the agent forward better than just a score, preventing unnecessary extra tool calls just to interpret the result.
- Evals in the online loop: The author's top priority scenario is having the agent run self-evaluations after drafting, fixing or flagging low groundedness immediately rather than relying on post-processing batches.
- Traces over scores: During troubleshooting, teams rely heavily on tool call traces rather than raw eval scores, as traces reveal exactly why an agent chose the wrong tool or entered a loop.
The author concludes with an open question: for state-altering tools, should gating mechanisms be applied before the call or after the response?
More from coding & agent
- A better path to agent autonomy is running waves, finding friction, and iterating — JnBrymn · 2026-07-22
- Coding agents are heading toward an AI-writes, AI-reviews, human-approves workflow — aftahi_ai · 2026-07-22
- oMLX 0.5.2 adds Mac menu-bar stats, low-bit decode kernels, and faster downloads — awnihannun · 2026-07-22
- GitHub review bot hits its PR limit and forces a 39-minute cooldown — DanielLockyer · 2026-07-22
- Max reasoning effort appears to be mobile-only in Codex Remote, not desktop — GabGarrett · 2026-07-22
- A Reddit demo argues online stores should expose carts and pricing through MCP — gelembjuk · 2026-07-22