A cloud platform exposes 46 MCP tools and asks how to gate destructive actions
Affectionate_Date749 · reddit · 2026-07-29
A cloud platform team exposes 46 MCP tools and asks how others gate destructive actions
A co-founder describes a cloud platform wired entirely to MCP, exposing 46 tools for databases, containers, functions, storage, auth, logs, and usage.
Design choices
- The server is stateless and uses plain POSTs only; a GET returns 405.
- Each agent gets its own scoped key rather than borrowing a human credential.
- Permissions are explicit grants.
- Consequential actions stop for one-tap human approval.
- Every call writes a receipt that the agent can read back to verify the result.
The open question
The team is unsure whether one-tap approval is enough friction, or whether risk tiers per action would be better despite the added configuration complexity.
The post is effectively a real-world discussion of how to safely expose powerful tool surfaces to agents.
More from coding & agent
- Kernel Forge uses MCTS to optimize CUDA kernels and beats PyTorch baselines on 14 cases — omarsar0 · 2026-07-29
- Every says Opus 5 works better when given one clear job and left alone — every · 2026-07-29
- Claude Code’s Fable xhigh wins trust after weeks of consistent performance — doodlestein · 2026-07-29
- Relume MCP and Claude Code power a Markdown-first design workflow — michalmalewicz · 2026-07-29
- A research talk compares SWE-bench, CodeClash, and ProgramBench for agentic coding — OfirPress · 2026-07-29
- Kuna debuts as an agent-driven Rust decompiler, ranking near Hex-Rays in structuring — moyix · 2026-07-29