Postman wrote itself into my coding agent's config. How to check yours, and what I want from API tools instead
At the end of October a Postman desktop update added Postman plugins to Claude Code and Codex on some machines, along with a hook that runs every time an agent session starts. Nobody was asked. After a GitHub issue hit t
At the end of October a Postman desktop update added Postman plugins to Claude Code and Codex on some machines, along with a hook that runs every time an agent session starts. Nobody was asked. After a GitHub issue hit the front page of Hacker News, Postman said it would revert the change.
I build an open-source API client, so I have an obvious bias here. But the cleanup part of this post is useful whatever you use, so I'll start with that.
What was reported
The original report was GitHub issue #14370 on postmanlabs/postman-app-support, opened on October 25. The issue has since been deleted, so everything below comes from people who read it at the time (explainx, byteiota, buildmvpfast):
-
~/.claude/settings.jsongot apostman@claude-plugins-officialplugin, enabled for the whole user, so it loaded in every project. -
~/.codex/config.tomlgot a Postman marketplace, an enabledpostman@postmanplugin, and a pre-approved hook entry. - A
SessionStarthook printed a note into the agent's context telling it to load Postman'sapi-engineerskill before any non-trivial API task. - A remote MCP server at
mcp.postman.comwas registered. - The Postman CLI appeared in
~/.local/bin/postman.
What was not shown: that any API data actually left anyone's machine. In the original report the remote MCP connection failed authentication. The problem is consent, not a confirmed leak, and I think that's bad enough.
Postman's response in the issue, as quoted by buildmvpfast, was that the change was meant "to simplify the setup" and that they are "reverting this change".
Check your machine
Claude Code:
claude plugin list | grep -i postman
grep -i postman ~/.claude/settings.json
grep -i postman ~/.claude/plugins/installed_plugins.json
Codex (and Copilot CLI, which one user also reported):
grep -n postman ~/.codex/config.toml ~/.copilot/settings.json 2>/dev/null
On macOS, the Postman log shows whether the auto-install ran:
grep -E "auto-existing-upgrade|autoInstall" ~/Library/Application\ Support/Postman/logs/main.log
Remove it
Postman's own remover, pinned to the latest version so an older cached copy doesn't run instead:
npx @postman/postman-plugin@latest remove
If you only want the Claude Code plugin gone:
claude plugin uninstall postman@claude-plugins-official --scope user
Then run the checks again, delete ~/.local/bin/postman if you don't use the CLI, and restart your agents. If something is still in ~/.codex/config.toml, the [marketplaces.postman] and [plugins."postman@postman"] blocks and the postman hook entry can be deleted by hand.
Why this one bothers me
A plugin with a SessionStart hook isn't a passive integration. It's text that lands in the context of every agent session you start, in every repo, and it steers what the agent does next. The approval prompt that Claude Code and Codex show for hooks is there exactly so that a human reads that text once. Writing the approval yourself skips the only check there is.
And the tool doing it is the one that holds your API keys. An API client is close to the most sensitive thing on a backend developer's laptop: staging tokens, production tokens, the cookie you copied from a browser at 2 a.m.
I still want my agent to use an API client. Saved requests with working auth are much better than an agent improvising curl commands with tokens pasted into the chat. I just want it on different terms.
What I'd ask of any API tool that talks to an agent
- Nothing changes in my agent's config until I run a command. Opening the app or updating it is not consent.
- It shows what it will change before it changes it, and there's one command that undoes it.
- It adds tools, not instructions. An MCP server describing its own tools is fine. A note telling the agent to prefer one vendor for all API work is not.
- Secrets never reach the model. Not in the request the agent reads, not in the response, not when the server echoes my token back in an error.
- Production is off limits unless I say otherwise, and a read-only mode exists for when I only want the agent to look.
How I ended up doing it in Kurlo
Here's how that list looks in Kurlo, the client I work on. It's MIT, no account, no telemetry, and collections are YAML files in your repo.
1β2. One explicit command, with a preview and an undo.
npx -y @kurlo/cli mcp install claude-code --env Local --dry-run
npx -y @kurlo/cli mcp install claude-code --env Local
npx -y @kurlo/cli mcp uninstall
--dry-run prints the change and touches nothing. Without it, Kurlo prints every file it changed and leaves the rest of the file alone. If your config file has comments in it, Kurlo doesn't edit it at all and prints the entry for you to paste.
When you use Claude Code from inside the Kurlo app, the tools are passed to that one session as command-line flags. Nothing is written to ~/.claude.
3. One MCP server entry. No plugin, no skill, no hook, nothing that runs at session start.
4. Secrets are used, but the model sees [secret]. Variables marked secret are filled in when the request is sent and masked in everything returned to the agent: the resolved URL, response headers and body, test messages and script logs. A server that echoes your token back doesn't hand it to the model. The write tools can't store a secret either, so the agent can't leak one into a committed YAML file.
5. Protected environments and read-only mode. Mark an environment "Protect from AI agents" and every call that would use it is refused. The agent is told to ask you to run it yourself. --read-only drops every tool that writes.
What you get for that: the agent can list your saved requests, run them with real auth and tests, explain why the last request you sent from the app failed, and write new requests and pm.test checks as reviewable YAML diffs. Postman collections import, and their scripts run as written.
The bigger point
This is going to keep happening. Every developer tool now wants to be the default inside your coding agent, and the agent's config is just a file in your home directory that any installer can write to. The cheapest defense I know: keep ~/.claude/settings.json and ~/.codex/config.toml in a dotfiles repo and look at git diff after installing or updating anything.
Did anyone else find Postman entries in their configs? And which other tools have you caught doing this?
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.