Claude Code Plugin Marketplace Governance: What Solo Founders Need to Know in 2026

By Rakshit Yadav (@yadavrakshit60)•Nov 2026•9 min read

Short answer: Claude Code plugins run as fully trusted code in your session the moment you /plugin install them — there's no binary signing for Claude plugins yet, so a malicious or compromised marketplace entry can register a hook script that executes on every tool call with the same permissions you have. For a solo founder, the practical governance move isn't an enterprise allowlist system; it's a 5-minute manual audit before each install, plus knowing when a managed-settings.json allowlist is actually worth setting up. Below is the checklist, the real risk Thoughtworks flagged, and where that risk doesn't apply if you're running AgenticKit's pre-reviewed agent team instead of installing plugins piecemeal.

Why plugins run fully trusted code in your session

A Claude Code plugin isn't a sandboxed add-on — it's a bundle of commands, skills, subagent definitions, and hook scripts that get registered directly into your session the moment you run /plugin install <name>@<marketplace>. As covered in the plugins guide, the distinction that matters is plugin vs. skill: a skill is inert markdown the agent opts into when its description matches your task, but a plugin's hook scripts run unconditionally on events like PreToolUse and PostToolUse — meaning arbitrary code executes on every matching tool call, with access to whatever the Claude Code process itself can touch (your filesystem, your shell, your environment variables).

That's a materially different trust boundary than installing an npm package in a sandboxed CI job. A plugin's hook runs inline, synchronously, in the same session where you might be pasting API keys into a prompt or running git push. If the hook is malicious or simply careless, it sees everything that session sees.

The supply-chain risk Thoughtworks flagged (no signing yet)

Thoughtworks' Technology Radar listed the Claude Code plugin marketplace as a tool to watch, flagging exactly this: plugins are distributed from git repos with no cryptographic signing or verified-publisher system as of their review (source: Thoughtworks Radar — Claude Code Plugin Marketplace). Anyone can stand up a marketplace repo, and nothing at the install layer verifies that the code in it matches what the README claims, or that the maintainer who published version 1.0 is the same one pushing version 1.3.

A separate community writeup made the same point from the publishing side: a marketplace is "more than a git repo" because without a manifest schema and review process, there's nothing stopping a plugin from being swapped out after people start trusting the name (source: Your Claude Plugin Marketplace Needs More Than a Git Repo). Star count and install count tell you popularity, not safety — neither is cryptographically tied to the code you're about to run.

The practical upshot for a solo founder: until Claude plugins ship binary signing, "official" vs. "community" marketplace is your main trust signal, not a package checksum. Anthropic's own anthropics/claude-code marketplace carries vendor-level review; anything else is you vouching for a stranger's repo.

Managed settings allowlists — do solo founders need them?

Claude Code supports a managed-settings.json file that an organization can use to lock down which plugins, MCP servers, or marketplaces a session is allowed to install from — enforced centrally, not just as a per-session setting a developer can override. This exists for the governance case Thoughtworks and others are pointing at: a team where you can't audit every engineer's plugin installs individually, so you push a managed allowlist instead.

If you're a solo founder working alone in your own repo, you're almost certainly over-scoped by setting one up. A managed-settings allowlist is infrastructure for enforcing policy across people you don't personally watch install things — if you're the only person running /plugin install, the policy and the person are the same, and the 5-minute checklist below does the same job with less setup.

The allowlist earns its keep the moment either of these becomes true:

  • You bring on a second developer (contractor, co-founder, early hire) who'll have their own Claude Code session touching the same repo.
  • You're running Claude Code in a CI or automation context where no human reviews each install before it happens.

Until then, skip the allowlist infrastructure and spend the time on the checklist instead — it catches the same class of problem with zero config files to maintain.

A 5-minute checklist before installing any plugin

  1. Read the marketplace repo's marketplace.json and the plugin's actual source, not the README. A plugin is usually a folder of .md files, a few commands, and optionally a hook script — five minutes is enough to read all of it.
  2. Grep specifically for hook registrations. Commands and skills are inert until the agent chooses to use them; a PreToolUse or PostToolUse hook runs on every matching call whether you intended to invoke it or not. If you can't find a hook script, you're not done auditing — check again before assuming there isn't one.
  3. Check commit history, not stars. A repo with no commits since its initial push, or a changelog that stopped updating months ago, is a plugin nobody is maintaining when the plugin schema changes under it.
  4. Confirm the publisher matches who you think it is. For anything touching credentials, deploys, or git history, prefer the official anthropics/claude-code marketplace or a maintainer whose other work you already trust — this matters more without signing in place, since there's no cryptographic proof tying a later update to the same author.
  5. Install, test once, then /plugin list and remove what you're not actively using. Every installed plugin with a hook is attack surface sitting in your session even when you're not thinking about it; uninstall discipline is the cheapest governance control you have.
# Before installing, always check what's already there and where it comes from
/plugin marketplace list
/plugin list

# Install from the marketplace you've vetted
/plugin install <plugin-name>@<marketplace-name>

# Periodically prune
/plugin uninstall <plugin-name>

This is the same discipline the plugins guide recommends per-install — the governance question here is really "how do I apply that checklist consistently without enterprise tooling," and the answer is: run it every time, manually, because at solo-founder scale that's faster than building the infrastructure to enforce it on yourself.

How AgenticKit's agent team avoids the trust problem

AgenticKit doesn't distribute itself as a plugin marketplace entry you install from a third party — it's a direct install that drops 34 subagents, 61 skills, and 47 commands into your repo as markdown and configuration you can read in full before (and after) you add it. None of those pieces register hook scripts that execute unconditionally; a subagent runs only when invoked for a specific task, the way Anthropic's own guidance on subagents describes using them (source: How and when to use subagents in Claude Code), and a skill stays inert until its description matches what you're asking for — the same skill-vs-plugin distinction that makes a skill a lower-trust artifact than a plugin with a hook.

That's not a signing solution — AgenticKit's content isn't cryptographically signed either, since nothing in the current Claude Code ecosystem is. But it sidesteps the specific risk this post is about: you're not assembling trust piecemeal across a dozen unrelated marketplace authors, each one a separate audit. It's one repo, one set of markdown files, reviewed once, with the security-auditor agent and code-reviewer agent available to apply the same scrutiny to your own code going forward.

FAQ

Are Claude Code plugins signed or verified? No. As of this writing, there's no binary signing or verified-publisher system for Claude plugins — Thoughtworks' Technology Radar flagged this explicitly as a supply-chain risk to watch. Trust currently comes down to which marketplace a plugin is published under and how carefully you read its source, not a checksum you can verify.

What's the actual risk of installing an unaudited plugin? A plugin can register a hook script (PreToolUse, PostToolUse) that runs on every matching tool call in your session, with the same access the Claude Code process has — your filesystem, shell, and environment. A skill or command alone carries less risk since those only execute when the agent chooses to use them for a matching task.

Do I need a managed-settings.json allowlist as a solo founder? Usually not. Allowlists solve the problem of enforcing policy across people you don't personally watch install things — if you're the only person running /plugin install in your repo, the 5-minute manual checklist above does the same job without the setup. Add one once you bring on a second developer or run Claude Code in unattended CI.

Is the official anthropics/claude-code marketplace safer than community marketplaces? It carries vendor-level review, which is the strongest trust signal currently available in the absence of signing. Community marketplaces can still be fine, but the audit checklist above (source read, hook check, commit history, publisher identity) matters more for those since nothing is independently verifying them.

How is AgenticKit different from installing a plugin? AgenticKit is a direct repo install of markdown-based agents, skills, and commands — not a marketplace-distributed plugin with hook scripts that execute unconditionally. See the plugins guide for the full plugin-vs-skill-vs-agent breakdown.

Next steps

If you're installing plugins today, run the 5-minute checklist above before every /plugin install, and default to the official anthropics/claude-code marketplace for anything touching credentials or git history. Skip the managed-settings.json allowlist until you have a second person's session to govern — it's solving a team problem, not a solo one. If you'd rather not assemble trust plugin-by-plugin at all, AgenticKit's 34 agents and 61 skills cover the same workflow ground as one audited install.

Related Tools & Agents

🛠️ Free Tool: security-scan🛠️ Free Tool: kit-picker🤖 Agent: security-auditor🤖 Agent: code-reviewer⚡ Command: /auditSkill: security-basics

Skip plugin-by-plugin trust audits

AgenticKit ships 34 agents, 61 skills, and 47 commands as a single reviewable install instead of a marketplace stack you'd have to audit one plugin at a time.

Browse agents →