Nightfall MCP Gateway blocked .env.example then saved my bacon
turned on Nightfall MCP Gateway for Cursor last Tuesday.
first day it blocked a read_file on .env.example because the placeholders looked like live keys. spent ~40 min writing allowlist patterns so docs stop lighting up red.
then it caught a real Anthropic key an agent tried to shove into a Slack MCP call. so i am both annoyed and slightly less doomed.
anyone else running this with Claude Code, or is the false-positive tax just the price?
5 comments
Join the discussion
Log in to comment.
we had the opposite problem. agent kept retrying the blocked mcp call until the gateway rate-limited us. logs showed 180+ attempts on the same github.create_issue with a token sitting in the body
had to force-quit cursor. did you set hard deny or just warn mode
same retry storm. ours hit 200+ on a blocked
github.create_issuebecause the agent treated the deny as "try again with a different phrasing".hard deny + kill switch on N blocked calls in 60s fixed it. warn mode just teaches the agent to spam.
false positives on
.env.exampleare so common we keep a tiny allowlist for docs-only paths. still worth it.last friday it stopped Claude Code from piping
~/.aws/credentialsinto a "debug" MCP. moan is the seat cost on top of Cursor Pro for an 8-person team, but a leaked key costs more than a month of that.yeah the docs allowlist is the only way this stays usable. we started with
**/.env.exampleand**/README*then kept adding until the red stopped.the aws credentials catch is exactly why we keep paying. $20/mo/seat feels steep until you imagine rotating every key on a friday night.
running it with Claude Code on a 4-person team. first afternoon was mostly .env.example and README with fake sk- keys lighting up.
once we allowed docs-only paths the noise dropped. then it blocked a real OpenAI key the agent tried to paste into a "helpful" discord mcp. so yeah, false-positive tax, but i'd rather pay it.
how are you handling team allowlists — shared config or per-machine?