10 Hermes agent skills worth installing today

If your agent reads five files to answer one question and never forgets them, your context fills with material you already used. Hermes agent skills attack that problem directly, from a reader-mode extractor to a repository index. This article lists what each of the ten does, who published it, and where the numbers come from.

What are hermes agent skills?

Hermes agent skills are reusable extensions that one creator runs on his agent, shipped as Claude Code skill files or MCP servers rather than as model changes. They split into three groups by effect: three cut token cost, five add a capability the model lacked, and two coordinate several agents on one job. Each lives in its own GitHub repository, so the name, license and install path come from that repository, not from the video.

The distinction matters because "skill" describes the effect, not the packaging. A markdown skill file loads once and then sits in your conversation for the rest of the session, which is cheap when the file is small and expensive when it is not. An MCP server or platform runs outside the conversation and returns only the result. Both can make your agent cheaper or smarter; only one of them costs you context space for the whole session.

The three cost-cutting hermes agent skills

Three of the ten skills reduce what the agent reads or writes. Defuddle Stripe web pages down to reader-mode text, Caveman compresses the agent's own output, and Codebase Memory replaces repeated file reads with graph queries. Only one of the three is a true skill file, so treat the group as a cost category rather than a packaging category.

The transcript gives the clearest numbers for Caveman and Codebase Memory, and both have provenance worth keeping separate.

Defuddle and Kaveman: cut what goes in and what comes out

Defuddle converts a web page to reader-mode Markdown before the agent reads it, removing navigation, footers, cookie banners and ads. Caveman is an ultra-compressed output mode whose repository claims a 65 percent reduction in output tokens. Both are skill files, and both matter because output tokens cost more than input tokens on most models.

Input waste compounds. If your agent visits a page at turn four of a 40-turn session, every token it read from that page is still in the conversation and gets resent on each remaining turn. Cutting the page at the source removes the waste everywhere it would have been re-sent. Output waste compounds the same way, and it is the more expensive side: what the agent writes becomes part of the conversation and is sent back with every turn after.

The creator's own demo of Caveman shows the same information delivered without the preamble, the recap, the summary of what it just said, and the offer to help with anything else. The compressed version is not worse output; it is the same answer with the ceremony removed.

Why Defuddle is the one to install first

Defuddle is the simplest skill on the list and the lowest-risk install. It does not touch your agent's reasoning, its tools or its files. It changes one thing: what the agent sees when it opens a URL. If your agent does any research at all, the creator's position is that this skill is not optional, because it pays for itself on the first page you read.

Codebase Memory: stop paying for the same file twice

Codebase Memory is technically an MCP server rather than a skill file, so there is no markdown file to install. It indexes an entire repository into a graph once, and after that the agent queries the map instead of reading files. The repository claims 99 percent fewer tokens for exploration across 158 languages.

The arithmetic behind that claim is the part worth internalizing. A 3,000 token file your agent reads at turn four of a 40-turn session does not cost you 3,000 tokens. It costs you 3,000 tokens 37 more times. Multiply that by every file in an import chain and the cost of exploration dwarfs the cost of the actual work.

The creator ran the same test independently and measured 95 percent fewer tokens used, close to the repository's 99 percent figure and still a large reduction. Treat 99 percent as the vendor's claim and 95 percent as one user's replication, not as a guarantee for your repository.

Five skills that add ability the model did not have

Five of the ten skills give the agent something it could not do before: Humanizer, Agent Reach, the marketing pack, Composio, and Open Montage. The first three install quickly, and the last two are platforms rather than skill files, so budget an afternoon rather than a minute for those.

The order below follows the creator's own install advice: judgment and access first, infrastructure later.

Humanizer: 29.6 kilobytes of AI-writing tells

Humanizer is 29.6 kilobytes, the largest file on the list, and the only one the creator is careful about installing. It was built from the Wikipedia page on signs of AI-generated writing, turned into a skill that Stripe those tells out of your agent's output and rewrites it in a more natural voice. It auto-updates when the source page updates.

The cost is real: once loaded, those 29.6 kilobytes sit in your conversation for the rest of the session. The creator's rule is therefore different from the rest of the list. Invoke it late, invoke it once, and if you are deep in a long session, do the writing pass in a fresh session. The payoff is narrow but concrete. If you publish anything your agent wrote, this is the last thing between you and getting clocked as automated.

Agent Reach: senses for an agent that runs on a server

Agent Reach gives the agent read access to social platforms it otherwise cannot reach. X is a paid API, and Reddit returns a 403 error the moment you hit it from a cloud IP, so an agent running anywhere except your laptop gets nothing back when you ask it to research a competitor. Agent Reach covers X, Reddit, GitHub and other networks with no API fees.

The feature that makes it durable is failover. Social platforms change their rules and block access paths constantly, and Agent Reach keeps a backup route map, so your agent does not notice the outage. The remaining limitation is the one you would expect: it depends on routes that platforms can close, which is exactly why the failover layer exists.

Marketing skills: 49 skills of judgment, not tools

Marketing skills is a single pack of 49 skills covering SEO, paid ads, conversion, copywriting, buyer psychology and more. The reason it earns space is that these are not tools your agent already had. Your agent can write. What it does not know is what a good landing page does differently from a bad one, why one subject line beats another, or whether three ad variants are actually testing anything.

If you sell anything, the creator's advice is to install this pack before any of the infrastructure skills on this list, because the return shows up in revenue rather than in token savings.

Composio: connect the tools your business already runs on

Composio connects the agent to Gmail, Sheets, Slack, Notion, HubSpot, Salesforce and more than a thousand other tools through a single install, with no API key juggling and no OAuth flow to reverse engineer. Like Codebase Memory, it is a platform rather than a skill file, so there is no markdown file to look for.

The problem it removes is the one that blocks most people: connecting one tool by hand means a console, an OAuth grant and a settings page, and every new tool means repeating all of it. Composio turns that per-tool work into one connection the agent can reuse.

Open Montage: the first open source agentic video system

Open Montage is an open source agentic video production system with 12 pipelines, over 100 tools and more than 700 skill and knowledge files in one install. The workflow starts when you paste in a video you like. The agent reads the transcript, analyzes pacing, scene structure, key frames and tone, then comes back with two or three concepts, the exact tool path for each, a cost estimate and a preview.

You approve the direction before it generates a frame. That negotiation step is the difference from other agentic video tools the creator has used, which typically start burning money and hand you the result at the end. Open Montage asks first, and spends second.

Two skills that turn one agent into a team

Two of the ten skills coordinate several agents on a single job rather than improving one agent's cost or ability. These are the skills that matter when the work is irreversible, such as shipping a feature, running a migration, or sending a proposal to a client.

The two coordination skills

The failure mode both skills address is agreement. Ask your agent a question and it answers. Tell it the answer looks wrong and it agrees with you. Then say it was right the first time and it agrees with that too. A single session can be the most agreeable liar you will ever work with, because nothing in it verifies the result before you ship it.

Oh My Cloth: decompose, assign, verify

Oh My Cloth fixes the agreement problem by turning one session into a coordinated team. It decomposes the task into subtasks, assigns specialist agents or external workers, runs them in parallel or as staged pipelines, and verifies the result instead of stopping at a half-finished answer. For anything irreversible, the creator describes it as the quality gate the agent has been missing.

The practical effect is that you stop being the only check on the output. Verification becomes a step in the pipeline rather than a favor you do at the end.

Superpowers: a skills framework with a methodology

Superpowers has over 260,000 stars on GitHub and is one of the largest repositories in AI. It is not a single skill. It is a skills framework with a methodology for how work gets broken down and executed, and the creator calls it the closest thing to an operating system for your agent that he has found.

One honest caveat from the creator: it takes an afternoon to understand properly. It is the largest project in the video and the one he says gets discussed least relative to its size.

The ten skills at a glance

Skill	What it changes	Key claim or figure	Packaging
Defuddle	What the agent reads from a URL	Cuts a page to reader-mode Markdown	Skill file
Caveman	What the agent writes back	65% fewer output tokens (repo claim)	Skill file
Codebase Memory	How the agent explores a repo	99% fewer exploration tokens across 158 languages (repo claim); 95% measured by the creator	MCP server
Humanizer	How the agent's writing reads	29.6 KB file, Stripe-writing tells	Skill file
Agent Reach	Access to social platforms	X, Reddit, GitHub and more with no API fees	Skill file
Marketing skills	Selling judgment	49 skills in one pack	Skill pack
Composio	Access to business tools	1,000+ tools, one install	Platform
Open Montage	Video production	12 pipelines, 100+ tools, 700+ skill and knowledge files	Platform
Oh My Cloth	Coordination and verification	Decomposes, assigns, verifies	Skill pack
Superpowers	Agent operating framework	260,000+ GitHub stars	Framework

How to install these yourself

Installing a Hermes agent skill is mostly a matter of knowing which packaging you got, because the install path differs. The steps below follow the creator's own advice on ordering.

1. Install Defuddle first. It changes only what the agent reads from a URL, it costs nothing to try, and it pays for itself on the first page you read.
2. Add Caveman for output compression. Output is the expensive side on most models, and compressed output also stops inflating every later turn in the conversation.
3. Add Codebase Memory as an MCP server, not a markdown file, if your agent spends turns exploring a repository. Index the repo once, then let the agent query the map.
4. Add the marketing pack before the infrastructure skills if you sell anything, because judgment about copy and landing pages is what your agent is missing rather than another tool.
5. Add Composio when connecting business tools by hand has become the bottleneck, and Agent Reach if your agent runs on a server and needs to read social platforms.
6. Handle Humanizer separately. Invoke it late, once, and in a fresh session if the writing pass is long.
7. Add Superpowers and a coordination skill only when work is irreversible and you want verification built into the pipeline.

Honorable mentions: two skills the creator did not count

Two skills the creator runs did not make the ten, because they are his own. Mission Control is his Hermes agent fleet dashboard: task dispatch, agent health and live cost tracking on one screen for the whole stack. Nova is his YouTube skill, which researches ideas, checks his performance data and drafts video ideas. Nova suggested the video this article is based on, and the creator credits it as the reason half the list exists, since every skill on it was tested against real work before it made the cut.

Frequently asked questions

• What are hermes agent skills in simple terms? They are reusable extensions installed into a Claude Code agent, shipped either as markdown skill files or as MCP servers. Skill files load into the agent's context and stay there for the session; MCP servers run outside it. The ten in this list fall into three groups: three reduce token cost, five add new abilities, and two coordinate multiple agents.
• Which hermes agent skill should I install first? Defuddle. It is the simplest install, it does not change the agent's reasoning or tools, and it cuts the page down to reader-mode Markdown before the agent reads it, which the creator says pays for itself on the first page you read.
• How much do these skills actually save? The Caveman repository claims a 65 percent reduction in output tokens. The Codebase Memory repository claims 99 percent fewer tokens for codebase exploration across 158 languages, and the creator measured 95 percent fewer tokens in his own test. Treat vendor numbers as vendor numbers.
• Why is Codebase Memory not a skill file? It is an MCP server. There is no markdown file to install, but the effect is the same: it indexes the repository into a graph once, and the agent queries that graph instead of reading files again.
• Can I install all ten at once? You can, but the creator's advice is to take the first three tonight and add the rest as the need appears. Skills are free to install; what costs you is burying the good ones under dozens you never use.
• Why does Humanizer come with a warning? It is 29.6 kilobytes, the largest file on the list, and once loaded it sits in your conversation for the rest of the session. Invoke it late, invoke it once, and consider a fresh session for the writing pass.
• What is the biggest project on this list? Superpowers, a skills framework with over 260,000 GitHub stars that the creator describes as the closest thing he has found to an operating system for an agent. It takes an afternoon to understand properly.
• Do any of these skills cost API fees? Agent Reach is explicitly built for zero API fees across X, Reddit, GitHub and other networks, since X is a paid API and Reddit blocks cloud IPs with a 403 error. Composio covers 1,000+ business tools without per-tool OAuth work.
• How often does this list change? The creator says the list changes every month, which is why the advice is to start with the three that cut cost today rather than trying to install everything at once.

Where to go deeper

The creator has published the source video for this list, and every repository mentioned is linked in its description. A related Portuguese-language channel, Dev Doido do canal do youtube, covers agent tooling in the same practical style, and crazystack.com.br collects developer tooling references of this kind.

The pattern across all ten skills is the same one the creator ends on: installing skills is free, and what costs you is the pile you never use. The measure of an agent setup is not how many extensions it has but how many of them are actually firing on real work.

The skills here solve two problems: your agent reads more than it needs, and it agrees with you instead of checking itself. If you have explained either problem out loud in a video, the explanation is already worth publishing. Paste the YouTube URL into Skalablog, let it transcribe the video, and generate an article from it.

Skala Blog

Source: 10 Hermes Agent Skills You NEED To Install Today