Postman Desktop Reportedly Added Claude Code and Codex Plugins Without Consent

A desktop app that can alter an AI coding agent’s settings raises a basic question of control: who gets to decide what instructions the agent follows? A developer says Postman’s macOS app installed integrations into Claude Code and OpenAI Codex without consent, then tried to mark its Codex startup hook as trusted.
The GitHub user “indicava” reported the behavior on October 25. The issue surfaced after Codex began returning authentication errors and repeatedly loading Postman skills. The report covers one Mac; the behavior hasn’t been independently reproduced on other systems.
Changes to both coding agents
The developer says launching Postman desktop version 12.33.4 triggered installations for both agents. Logs included in the report record an event on October 24 labeled “auto-existing-upgrade,” with results showing plugins installed for Claude Code and Codex.
The Codex user configuration at ~/.codex/config.toml had also been changed: it listed an enabled Postman plugin, a new plugin marketplace, and a trusted session-start hook. Claude Code’s configuration showed a Postman plugin, while a Postman CLI executable appeared on the system with timestamps matching the installation.
Those artifacts describe changes to local agent configuration and a connection to Postman’s remote MCP server at https://mcp.postman.com/mcp. Codex repeatedly failed to authenticate to that server, according to the report. The evidence presented doesn’t show that Postman collected or transmitted API data.
What the hook approval does
The reported installation also raises a question about how integrations gain the agent’s trust. Postman’s documentation says its installer can approve a Codex session-start hook automatically. Someone installing the plugin from within Codex must review and approve the hook themselves.
The installer source describes the process: it checks whether Codex already trusts the hook and, if not, uses config/batchWrite to add its trusted hash. The installer verifies that the hook hasn’t changed. If it can’t confirm approval, it tells the developer to review the hook in Codex.
A session-start hook can give an agent instructions as it starts, before the developer makes a specific request. Postman’s startup instructions tell Codex to load the postman:api-engineer skill for most API development tasks.
Codex is still expected to prioritize the developer’s instructions. But a trusted hook can make Postman’s tools a default part of how it handles API work. That may suit someone who chose the integration; the concern in this report is that Postman’s desktop app allegedly added it to an existing environment without asking.
Claims and evidence diverge
The report drew criticism on GitHub and X. Security researcher Frank Denis called the integrations “junk” and criticized the installer’s attempt to trust its own startup hook.
Siddharth Garimella claimed Postman’s hooks intercept API data and send it to the company. The GitHub report offers no evidence for that claim. It documents agent-setting changes and an MCP connection, rather than data collection or transmission.
One GitHub commenter called the behavior “unacceptable, even creating risk in corporate environment.” Another urged Postman to explain the behavior and seek permission.
Consent and cleanup remain open
The developer is asking Postman to stop installing integrations into existing agent environments without permission, leave hook approval to Codex, and provide a complete way to reverse the changes. Postman documents an uninstall command, but the developer says removal should also clear agent settings, trusted hooks, MCP connections, and installed CLI tools.
Until the behavior is clarified, developers can inspect Claude Code and Codex settings for plugins, hooks, and MCP connections they don’t recognize. The complaint joins wider calls for more developer control over coding agents, including recent criticism of agent refusals.
Postman’s installer can approve its own Codex hook without the usual manual review.
The unresolved issue is why opening an existing Postman installation reportedly initiated these changes without the user’s permission. The New Stack asked Postman about automatic installation, hook approval, and consent; the company had not responded by publication.
The post A developer opened Postman. Then two AI coding agents changed without their permission. appeared first on The New Stack.