Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Audit apps that installed agent plugins for you

Written November 5, 2026

Postman Desktop installed Codex and Claude plugins without asking. Check what it added, what it can reach, and how to remove it.

What happened

On October 24, a developer opened Postman Desktop 12.33.4 on a Mac. Postman's own log then recorded an automatic install: {"source":"auto-existing-upgrade","results":["claude-code:installed","codex:installed"]}. The developer filed the report as GitHub issue #14370 on October 25, with log lines, config excerpts, and pinned plugin source. The install ran as an upgrade path for an existing desktop user, not as a step the user clicked through. A later commenter in the same thread reported the same auto-install on a desktop app that updated from 12.31.3 to 12.35.1, so the behaviour is not pinned to one build.

The report is careful about what it does and does not show. It establishes unwanted integration and instruction changes. It does not claim data left the machine: the remote MCP connection the install added failed with an authentication error, and the sequence has not been reproduced on a clean Mac. Treat anything beyond that as unconfirmed.

What it added

In Codex, the install changed the global config file ~/.codex/config.toml. It added a [marketplaces.postman] section pointing at Postman's plugin repository, enabled a postman@postman plugin, and wrote a trust record under [hooks.state] for the plugin's startup hook. It also registered a SessionStart hook, matched on startup|clear|compact, that prints an instruction file telling the agent: before responding to any non-trivial API engineering task — designing, implementing, mocking, testing, monitoring, documenting, or deploying an API — load the postman:api-engineer skill and follow it. That instruction is far broader than a user asking for Postman on one task, and it runs at the start of every session.

The hook's trust record is the sharpest part. Codex normally asks the user to review and trust a hook definition before it can run. The installer's own source contains a function, approvePostmanHook, that reads the hook hash and writes the trusted_hash into Codex's trust record itself, with the source comment "Approves only Postman's current hook after installation". The report notes the trust record alone does not prove who created that specific record on the machine, so the code shows the approval attempt, not proof of a particular interaction.

In Claude Code, the install enabled postman@claude-plugins-official version 2.0.1 in ~/.claude/settings.json, recorded in ~/.claude/plugins/installed_plugins.json with timestamps matching the install window. It also placed a Postman executable in ~/.local/bin, and pointed a Codex MCP config at https://mcp.postman.com/mcp. Codex logs show that connection failing with "Auth required" before any session completed.

Audit steps

Run these checks on any machine where Postman Desktop has been opened. Do them on a clean machine if you want to see what the install does without touching an affected setup.

  1. Check the Codex config. Open ~/.codex/config.toml and look for a [marketplaces.postman] section, a [plugins."postman@postman"] entry with enabled = true, and any [hooks.state] record naming Postman. Any of the three means the install reached Codex.
  2. Check the Claude Code config. Open ~/.claude/settings.json and look for postman@claude-plugins-official. Then check ~/.claude/plugins/installed_plugins.json for the same entry and its install timestamp.
  3. Check for the CLI. List ~/.local/bin and look for a Postman executable with a modification time near the Postman install or update.
  4. Check the MCP connection. Look for a Postman MCP entry pointing at https://mcp.postman.com/mcp in the Codex config, and check Codex logs for MCP authentication failures naming Postman.
  5. Check what the hook tells the agent. If a Postman hook is registered, read the session-start instruction it prints. If it routes broad API work through a Postman skill without you asking, that is the instruction to remove, not just the plugin entry.

Removal, and what it does not undo

Remove the entries by hand: delete the Postman marketplace and plugin sections from ~/.codex/config.toml, delete the Postman hook trust record from [hooks.state], remove the Postman entry from ~/.claude/settings.json and from installed_plugins.json, remove the Postman executable from ~/.local/bin, and remove the Postman MCP entry. Then restart the agent and check every location above again.

Two caveats. First, the removal steps have not been tested on an affected machine, and the hook trust record may survive the usual remover — checking it again after removal is part of the job, not an optional extra. Second, no complete rollback path has been published: the issue asks Postman for one covering agent configs, hook trust, marketplaces, caches, MCP connections, and the CLI, with removal respected rather than reinstalled on the next update. Until such a path exists, check again after the next Postman update.

What Postman has said

The two accounts disagree, so neither is treated here as settled. No maintainer reply appeared in the GitHub thread itself as the early coverage read it. Separately, Postman's head of product was reported on November 3 as acknowledging the issue and committing to clearer opt-in mechanics for existing users, in a statement delivered through a press agency. The thread asks Postman to stop automatic agent-plugin installation for existing desktop users pending explicit opt-in, remove installer-side hook self-approval, and narrow the startup instruction to explicit Postman requests. Whether the clearer opt-in arrives, and what it covers, is still open.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Check a Cursor machine for the tampered Claude Code extension · Audit a repository's git config before any agent opens it · Check your Codex and Claude Code config for relay tampering