How to choose the best Claude Code agents on aitmpl (Claude Code Templates) without installing a risky one

aitmpl.com (Claude Code Templates) is the largest catalog of Claude Code agents: more than 400 subagents you install with a single npx, covering frontend, backend, security, DevOps and marketing. The hard part is choosing. Many look alike, almost all claim to be "experts", and each one lands in your project with permission to read, write and sometimes run commands. In this article we downloaded the data for the whole catalog, re-ran the NVIDIA SkillSpector scans on our own machine, and put together a six-point method to tell which aitmpl agent is best and which one to skip.

If you are not sure yet how a subagent differs from a skill, start with the Claude Code skills guide. For a deep dive on the scanner, read the SkillSpector with find-skills article.

What aitmpl.com is and what a Claude Code agent is

Claude Code Templates is an open source project by Daniel Ávila at github.com/davila7/claude-code-templates, MIT licensed. It has about 33,100 stars and 3,900 forks, and the CLI version we reviewed is 1.33.1 (commit 30f72e1, November 11, 2026). The aitmpl.com site is its storefront: agents, skills, commands, hooks, MCP servers, settings, mods and plugins, each with its own page and install command.

An agent (or subagent) is a Markdown file with YAML frontmatter that Claude Code loads from .claude/agents/. Claude delegates tasks to it with its own context, its own system prompt and, if set, its own tools and model. This is the top of code-reviewer, the second most downloaded:

---
name: code-reviewer
description: "Use this agent when you need to conduct comprehensive code reviews..."
tools: Read, Write, Edit, Bash, Glob, Grep
---

Three lines decide almost everything: description (when Claude calls it), tools (what it can do) and model (which model it runs on and what it costs).

Install it like this:

npx claude-code-templates@latest --agent development-tools/code-reviewer

The catalog in numbers

To avoid guessing we pulled docs/components.json from the repository, the file the site reads. It lists every agent with its category, downloads and SkillSpector result:

The site advertises "600+ AI Agents"; the repository holds 423 agent files in its current version. The gap may come from external agents or how they are counted, but what matters when choosing is the specific file, not the total.

The ten most downloaded, with their results:

What "Scanned by NVIDIA SkillSpector" means

Every aitmpl component carries a Scanned by NVIDIA SkillSpector seal. It is a real plus: the catalog runs every agent through NVIDIA's security scanner (Apache-2.0) and publishes the findings on the component page, which almost no other agent directory does. But read it correctly. The site's own code says so in a comment on the seal component:

"Neutral on purpose: it says the component went through the scan, not how risky it is."

In other words, the seal means "scanned", not "safe". The three agents marked DO_NOT_INSTALL carry it too. What matters is behind it: the score (0-100), the recommendation and the list of findings.

How it is produced, according to scripts/skillspector_catalog.py and the project's security guide:

• An agent is a single file, so it is copied into a temp folder as SKILL.md and scanned with skillspector scan <folder> --no-llm --format json.
• It is static analysis only (patterns, YARA, references). The LLM phase, which confirms or dismisses findings, is not used.
• The pinned version is SkillSpector 2.12.0, and an agent is rescanned only when its content changes.
• In pull requests, a skill scoring above 50 is blocked. For agents, commands, hooks and MCPs the report is posted on the PR without blocking, because it raises more false positives.
• The signal is display-only: it does not change ordering or search in the catalog. Agents at the top are the most downloaded, not the safest.

We re-ran the scan ourselves

We installed the same version and scanned three agents the way the catalog does:

uv tool install 'skillspector @ git+https://github.com/NVIDIA/skillspector.git@v2.12.0'
mkdir -p scan/mcp-testing-engineer
cp .claude/agents/mcp-testing-engineer.md scan/mcp-testing-engineer/SKILL.md
skillspector scan scan/mcp-testing-engineer --no-llm --format json

Each scan took about 5 seconds and the results matched the site exactly: code-reviewer 0/SAFE, ui-ux-designer 29/CAUTION and mcp-testing-engineer 87/CRITICAL/DO_NOT_INSTALL. The catalog data is reproducible, which already counts for a lot.

Six criteria for choosing an aitmpl agent

1. It does a concrete task you repeat

The best agent is not the most "senior" one but the one that fits a task you do every week. Read the full description: Claude uses it to decide when to delegate. If it says "Use PROACTIVELY", the agent will trigger on its own often; if you don't want that, edit the sentence after installing. Avoid installing 20 agents "just in case": every description competes for Claude's attention.

2. Downloads: good for ruling out, not for choosing

Downloads tell you what more people have tried, and in a 423-agent catalog they help you filter. But the number measures neither quality nor safety: command-expert has 2,047 downloads and is one of the three DO_NOT_INSTALL, and search-specialist, with 7,028, scores 49, two points short of HIGH. Use downloads for the shortlist and nothing more.

3. Tools: the most overlooked criterion

The tools field is the agent's real permission set. If it is missing, the agent inherits every tool in the session, including the MCP servers you have connected. Ten catalog agents are like that.

Rule of thumb: it should ask for just what its job needs.

• A code reviewer needs to read (Read, Glob, Grep). code-reviewer also asks for Write, Edit, Bash; if you only want its opinion, remove them after installing.
• search-specialist only asks for WebSearch, WebFetch: even with a high score, it cannot touch your files.
• An agent with Bash can run any command your Claude Code permissions don't block. 232 agents ask for it.

4. The SkillSpector score, on its scale

Also check the analysis field. frontend-developer, the most downloaded, comes out CAUTION with 0 points and 0 findings. Why? Because its analysis is partial: the text mentions eslint.config.js and SkillSpector can't find that file next to the agent. That happens to 162 agents. A CAUTION with 0 points and partial analysis is almost always that.

5. Read the findings: there are false positives (and real hits)

The scan is static with no LLM, so it flags words, not intent. We looked at the exact lines behind the worst scores:

The most common finding in the catalog (82 agents) is EA5 "External Model or Provider Selection", and it fires only because the frontmatter says model: sonnet. It is not dangerous, but it has a cost: that agent runs on Sonnet even if your session uses another model. The unpinned npx warnings ("rug pull" risk) are worth taking seriously.

Bottom line: the score sorts, but the decision comes from reading the flagged lines. For a semantic second opinion, scan it yourself with the LLM phase on, as shown in the SkillSpector article.

6. Scan what you installed, not what you saw

This is the least known point. The project's security guide says external skills and plugins are pinned to a commit, so what you install is what was scanned. That is not the case for the repository's own agents: the CLI downloads them at install time from the main branch:

githubUrl = `https://raw.githubusercontent.com/davila7/claude-code-templates/main/cli-tool/components/agents/${agentName}.md`;

If the file changed after the last scan, you install a version nobody has scanned yet. The fix takes 5 seconds: scan the installed file.

npx claude-code-templates@latest --agent development-tools/code-reviewer --yes
mkdir -p /tmp/scan/code-reviewer
cp .claude/agents/code-reviewer.md /tmp/scan/code-reviewer/SKILL.md
skillspector scan /tmp/scan/code-reviewer --no-llm

And commit .claude/agents/ to your repository so any future change shows up in a diff.

Dependencies between agents

Some agents assume others exist. frontend-developer starts with "Always begin by requesting project context from the context-manager", which is another catalog agent (10,837 downloads). If you don't install it, Claude will look for it and not find it. Before installing, search the text for other agent names and install only the ones you will use.

What it costs

• aitmpl.com and the CLI: free, MIT license. You can copy and modify the agents in commercial projects as long as you keep the license notice.
• SkillSpector in static mode (--no-llm): free, Apache-2.0, runs on your machine in seconds.
• SkillSpector with an LLM: you pay the tokens of the provider you configure; for one agent file it is a single call.
• The agent in use: cost is set by the model field. An agent with model: haiku is much cheaper than one with sonnet for mechanical tasks (formatting, searching, summarizing). If the agent sets no model, it uses your session's. Changing that line is the cheapest way to save.

Common mistakes

• Trusting the seal without opening the report. The seal only says a scan happened; the score and findings are on the component page.
• Rejecting an agent for a 0-point CAUTION. Check whether the analysis is partial because of a config file reference.
• Installing the reviewer with write access. Remove Write, Edit and Bash if you only want a report.
• Installing a whole category at once. More agents means more descriptions competing and more attack surface.
• Not versioning .claude/agents/. Without a commit you can't see what changed on the next reinstall.

Frequently asked questions

Are aitmpl agents safe?

Most are, according to SkillSpector: 254 of 423 come out SAFE and only 3 come out DO_NOT_INSTALL, two of them false positives we reviewed. Still, "safe" depends on the permissions you grant: limit tools and review the file before using it.

Does the "Scanned by NVIDIA SkillSpector" seal mean NVIDIA approves it?

No. It means the catalog ran the component through SkillSpector, NVIDIA's tool. It is not an NVIDIA certification nor a safety verdict; the site's own code calls it a seal that is "neutral on purpose".

Which aitmpl agent should I install first?

To start: code-reviewer (SAFE, 0 points) trimmed to read-only tools, and debugger (SAFE, 9 points) if you debug often. Both are among the most used and have clean reports.

Can I use the agents in client projects?

Yes. The repository is MIT: you can use and modify them, commercially too, keeping the copyright notice.

What is the difference between an agent and a skill?

A skill adds instructions and scripts that Claude loads when the task fits, inside the same conversation. An agent is a separate assistant with its own context, tools and model, to which Claude delegates a task. aitmpl offers both.

Conclusion

To choose well on aitmpl.com: start from a task you repeat, use downloads only for the shortlist, trim the tools, read the SkillSpector score on its scale, read the findings, and scan the file you actually installed. NVIDIA's seal is a great starting point because it gives you the report for free; the decision is still yours. Next, set up the full search-and-scan flow with find-skills and SkillSpector, and to organize the agent's work after installing it, see Superpowers.