MCP users debate field-level redaction instead of all-or-nothing tool access
Sad_Cover9067 · reddit · 2026-07-24
An MCP output could be sanitized field-by-field instead of fully allowed or blocked
The post raises a permission-fatigue problem in agent workflows: after enough prompts, the user may stop making meaningful security decisions and just click Allow.
The author suggests a middle ground for MCP tools. If a tool returns:
json
{
"accountid": "12345",
"subscription": "pro",
"passportnumber": "AB1234567"
}
the agent may only need the account ID and subscription status. Instead of forcing a binary choice—either expose everything or block the whole response—the tool could preserve structure while replacing sensitive fields locally:
json
{
"accountid": "12345",
"subscription": "pro",
"passportnumber": "{{SEALED1}}"
}
The idea is that the model can keep working with the redacted response, while the real value stays local and could be restored later by the user.
The author says they are experimenting with this in a local tool, but they are unsure where the control should live:
- inside the MCP server,
- in the client,
- or in a proxy layer between them.
They also worry that mutating tool output may introduce ambiguity for the agent.
More from coding & agent
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- Dev ships 1000+ commits with agents by pushing straight to main, no complex orchestration — Dimillian · 2026-10-27
- AI code editor Superset rebrands as Paradev, raises $11.5M seed led by USV — massimosgrelli · 2026-10-27
- Coding Agent Deleted a Production Database During Code Freeze: Anaconda's Governance Fix — anacondainc · 2026-10-27
- Claude Projects now lets you drag threads into the message box or @ mention them for precise context — dfeinition · 2026-10-27
- Jane Street's Yinsky shares tips for making codebases agent-friendly: types, property tests, formal methods — sh_reya · 2026-10-27