知Yizhi Daily
← Back to this edition

Conversations

Users report Postman desktop automatically installing Codex and Claude plugins and rewriting proxy configuration

Edition 2026-11-04 · 08:00 Beijing time

A problem report on Hacker News says that after the user opened Postman for Mac, it automatically installed plugins for Codex and Claude Code, remote MCP connections, startup hooks, and a CLI. The reporter argues this creates serious consent and trust-boundary issues, but also clearly states that the evidence does not prove data was exfiltrated.

The reporter says that after opening an existing Postman desktop installation some time later, coding agents began repeatedly showing MCP authentication warnings, and Codex repeatedly routed API-related tasks to Postman skills. After checking global user configuration, they found a user-level Postman plugin, startup-command hooks, a remote MCP connection, a Claude Code plugin, and a separately installed Postman CLI.

The listed hooks/codex-hooks.json registers a SessionStart command and outputs context when a session starts, is cleared, or is compacted, requiring the agent to load the postman:api-engineer skill before handling non-trivial tasks such as designing, implementing, simulating, testing, monitoring, documenting, or deploying APIs. The reporter argues that this instruction’s scope goes far beyond the user explicitly asking to use Postman.

The report also says the installer source code calls approvePostmanHook after installing the plugin, obtains the hook hash through hooks/list, and then writes trusted_hash using config/batchWrite. A source-code comment says it only approves Postman’s current hooks. The reporter also notes that the trust record itself cannot prove who performed a particular user interaction.

On the Claude Code side, ~/.claude/settings.json enabled postman@claude-plugins-official, and installation records show version 2.0.1 with installation and update times on October 24, 2026; a Postman executable also appeared in ~/.local/bin. The reporter emphasizes that the report demonstrates unwanted integrations and instruction changes, but does not claim to have proven that data was sent externally.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader