Hermes Agent for SEO: hands-on guide

Hermes Agent is the open-source AI agent framework from Nous Research that crossed 150,000 GitHub stars within three months of release. The premise: an agent that lives across your tools, builds skills from experience, schedules itself, talks to you on Telegram or Slack, and remembers what you have asked it before.

It is the most popular general-purpose agent framework of 2026 and SEO teams are starting to deploy it for genuine work. This guide covers what it does, what it costs, where it earns its keep on an SEO programme, where it will quietly hurt your rankings if you let it run unsupervised, and what a consultant adds on top.

What Hermes Agent actually is#

Strip away the marketing language and Hermes is three things bundled together:

1. A terminal UI with persistent memory. Type hermes and you get a chat with a learning loop. The agent stores conversations in a local FTS5 database, summarises old sessions on demand, and uses a dialectic user-modelling library called Honcho to build a deepening picture of who you are over weeks.
2. A gateway that connects to messaging platforms. Same agent, reachable from Telegram, Discord, Slack, WhatsApp, Signal, email and a web dashboard. One conversation can drift between platforms without losing context.
3. A skill and tool system. Hermes can spawn isolated subagents, run shell commands across seven execution backends (local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox), and write its own skills after solving complex tasks.

The model is not built in. You point Hermes at any provider - Nous Portal, OpenRouter with 200+ models, NVIDIA NIM, OpenAI, Anthropic, your own endpoint - and switch with hermes model whenever you want.

The combination is what makes it different from chatting with ChatGPT or running a single-shot Claude script. Hermes is a persistent operator that keeps working when you close the laptop.

Why SEO teams are paying attention#

Most SEO work is repeated, semi-structured, multi-tool, and ripe for an agent that does not need babysitting. A short list of jobs that map cleanly to Hermes:

• Pull yesterday's Search Console performance, flag the queries with double-digit position drops, and ping the on-call SEO on Slack
• Crawl the staging environment overnight, diff the output against last week, and report any new noindex tags or broken internal links
• Watch a competitor's sitemap for new URLs, summarise the content and notify the content team within the hour
• Generate technical SEO audit briefs that pull live Ahrefs data, GSC performance and GA4 conversion signals into a single report
• Schedule a Monday morning summary that combines all of the above into a 200-word brief sent over WhatsApp

You can hire an agency to do those things. You can also write Python cron jobs that do them. Hermes sits between those two options: cheaper than the agency, more natural-language and easier to modify than the cron jobs.

That is the appeal. The question for SEO teams is whether the appeal survives contact with the messy reality of search work.

The setup most teams get wrong#

The fastest installs I see in client work look like this:

# Linux, macOS or WSL2 - one-liner from the official docs
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.bashrc
hermes

Two minutes later you are in a working agent. The mistake is what comes next. People wire Hermes directly into their live production environment - the real CMS, the real schema deployment, the real internal link graph - because they trust the demo they just saw.

The agent then does something useful nine times out of ten and something catastrophic the tenth time. A skill that was tested on three URLs gets applied to three thousand. A schema generator hallucinates a Person property and pushes it live. A scheduled cleanup script deletes draft posts that the editor was about to publish.

The right setup splits the work into three tiers:

Hermes makes this easy if you remember to set it up. Use the command_approval setting to require human confirmation for any destructive command. Run write-capable skills inside a Docker or Daytona sandbox so a bad agent decision touches a container instead of your prod database.

Without those guardrails, the first time a model update changes how the agent reasons about a prompt - which happens every couple of months as OpenAI, Anthropic and Nous ship new versions - you will discover a regression in production rather than in staging.

SEO workflows where Hermes genuinely earns its keep#

These are the patterns I have seen work in real client engagements over the last three months.

Scheduled SERP and ranking monitoring#

The cron scheduler is built in. A natural-language instruction like "Every weekday at 9am, pull the top 50 ranking queries from Search Console for the last 7 days. Compare to the previous 7 days. If any query has dropped more than 10 positions, send a Slack alert to #seo-ops with the query, the URL, the old position and the new position. Skip queries with fewer than 10 impressions." compiles to a recurring job that runs unattended.

That replaces a manual check that no in-house SEO actually does reliably. The cost is one paid API call to the LLM plus the GSC API (free under quota).

Multi-tool orchestration in a single conversation#

This is the standout. In one Hermes session you can ask:

Pull the slow-loading pages from CrUX, cross-reference them with the pages converting well in GA4, look up which of those pages also appear in the top 20 for a high-volume query in Ahrefs, and rank-order the result by conversion-weighted opportunity.

Hermes calls GSC, GA4, Ahrefs and CrUX through MCP servers or REST APIs, then composes the answer. You skip the spreadsheet shuffle between four browser tabs.

The work that an SEO consultant does to prepare this workflow is non-trivial. The credentials wiring, the rate-limit handling, the schema mapping between tools and the decision rules about what counts as opportunity all need expert judgement. Hermes is the runtime. The setup is craft.

Skills that compound over time#

The skill system is the killer feature. After Hermes completes a complex task it can write itself a reusable skill - effectively a stored procedure with prompts, tool calls and decision logic - that becomes available as a slash command next time.

A skill called /serp-audit might:

1. Take a URL and a target query
2. Pull the live SERP via SerpAPI
3. Extract the top 10 results
4. Compare your page's headings, entities and answer blocks against the winners
5. Output a recommendations list scored by impact

Build that skill once with an expert in the room. Run it 200 times over the next year on different URLs. The marginal cost approaches zero.

The community Skills Hub (compatible with the agentskills.io open standard) already has SEO-flavoured contributions, but most need to be adapted to your stack before they are useful. The vibe is the same as installing a plugin from a marketplace - you treat the listed code as a starting point, not a finished tool.

Cross-platform reporting#

Most SEO reports get ignored. They land in an inbox at 7am on Monday morning, the recipient skims the summary, and the underlying detail never gets opened.

Hermes lets you deliver the same report through whatever channel a stakeholder actually checks. The CMO who reads WhatsApp gets a 200-word brief. The dev lead who lives in Slack gets a ticket-shaped summary with proposed fixes. The CEO who skim-reads email gets a one-paragraph headline. Same source data, different framing per channel, generated by the same agent in one cron run.

SEO workflows where Hermes is a footgun#

Three patterns I see attempted that go wrong without expert oversight.

Content generation at scale#

The temptation is obvious. Hermes plus a frontier model can draft a 1,500-word blog post in 90 seconds. Two clicks to publish. Done.

This is the exact pattern Google's Helpful Content system is trained to detect. Not because AI wrote the post - HCU does not care who wrote it - but because the prompt that fits in a single Hermes turn produces prose that lacks specific examples, real numbers, named sources, contradictions, opinions and the small dose of personality that earns trust.

I have seen sites publish 200 AI-drafted posts and watch their organic traffic halve in the following core update. The cost of the prompts was £8. The cost of the recovery was nine months of careful republishing.

Hermes does not solve the content quality problem. It accelerates whichever direction you point it. With a consultant in the loop, the workflow becomes: Hermes drafts the structure, the consultant writes the parts that need experience, the editor approves, the article ships. Volume drops to two posts a week. Traffic compounds.

Schema markup generation#

Schema is structured trust. Get it wrong and you either get nothing (the engines ignore broken markup) or you get penalised for misrepresentation (claiming a Review rating you do not have).

LLMs hallucinate schema fields. They invent Organization properties that do not exist. They use deprecated values for MedicalCondition. They populate Review.itemReviewed with the wrong entity type. They mark up content as HowTo when it is Article and Google quietly stops showing the rich result.

Hermes will generate confident-looking JSON-LD that validates in the schema generator but fails in real-world surfaces. The fix is not "use a better model" - the fix is having someone in the loop who knows that Service requires provider and areaServed to be properly nested, knows when LocalBusiness is appropriate and when it is overreach, knows that aggregateRating requires actual reviews to back it up, and reviews everything the agent proposes before it ships.

Agentic crawling on sites you do not own#

Hermes can crawl any URL you point it at. Several skills in the community hub do exactly this for competitor analysis.

The problem: a SiteGround or Cloudflare WAF will block a crawler that requests 50 URLs in 30 seconds from a residential IP. Your VPS gets banned. Worse, if the WAF logs include your domain in the referrer header (some do), you have just told a competitor that you are watching them.

A consultant who has run crawls professionally knows to rate-limit, rotate UA strings, respect robots.txt, never crawl behind authentication that does not belong to the agent, and avoid crawling competitor login or checkout pages even when robots permits it. Hermes will do whatever you ask. The judgement is yours.

Real costs, broken down#

The biggest cost driver is which model you use. Default to the smallest model that gets the job done. Use a frontier model only for the work that needs reasoning.

The setup work is the hidden cost. Wiring Hermes into a working SEO stack with credentials, MCP servers, scheduled jobs, skill prompts, output formatting and approval flows is a 2-5 day piece of expert work. Once it is built it largely runs itself, but the build is not optional.

Where an SEO consultant earns the cost#

Reading the above, you can probably see where the line is. Hermes is the runtime. Someone has to:

• Decide which questions are worth automating and which are not. Most clients want to automate the work that is hardest to automate and easiest to do manually.
• Build the prompts and skills with current SEO judgement baked in. The same task prompt that worked in 2024 misfires in 2026 because algorithm sensitivity has shifted.
• Set guardrails that survive a model upgrade. An agent that worked perfectly on Claude Sonnet 4.5 will sometimes do dangerous things on Sonnet 4.6 without warning. Tested behaviours need re-testing.
• Review what Hermes is about to publish, push or send. Particularly for schema, redirects, content and outreach. A wrong sitemap can deindex a site overnight.
• Hold the strategy. Hermes will happily optimise a metric. The hard part is knowing which metric to optimise this quarter and why.

If you already have those skills in-house, Hermes accelerates you. If you do not, Hermes builds a system that fails confidently and quietly.

That second failure mode is the one I see most often. The site does not crash. Rankings just slowly drift down because the agent kept doing the wrong thing for three months and nobody knew because the dashboard looked green.

Setup walkthrough (abbreviated)#

The full install is well-documented at hermes-agent.nousresearch.com/docs. The condensed version for an SEO context:

1. Install on a server, not your laptop. Hermes is built to live on a remote machine and have you talk to it from anywhere. A £5 VPS is enough.
2. Configure the model. hermes model and pick a provider. OpenRouter is the most flexible because it lets you swap models without changing keys.
3. Wire SEO credentials. GSC OAuth, GA4 service account, Ahrefs API token, SerpAPI key. Store them in the encrypted Hermes vault, not in env vars in plain text.
4. Set up MCP servers for the tools you use. Many SEO tools already have community MCP servers. If not, a short Python wrapper around their REST API does the job.
5. Enable command approval. hermes config set tools.command_approval true so the agent has to ask before running anything destructive.
6. Start with read-only skills. Add five or six skills that only read data and report. Run them for a fortnight. Watch for drift.
7. Add staged writes once read-only is boring. Schema proposals, content drafts, redirect maps - all with a human approval step.
8. Schedule the recurring jobs. Daily SERP check, weekly competitor diff, monthly schema audit.
9. Connect the gateway to whichever messaging platforms you actually use. Telegram is the smoothest, Slack is the most professional.
10. Set up logging. Either via Hermes' built-in session FTS5 search, or pipe to a real observability stack if you want traces.

Plan on a fortnight of setup and tweaking before you trust the agent with anything important. The investment is real but the running cost after is trivial.

Alternatives worth considering#

Hermes is not the only agent framework. The 2026 shortlist for SEO use:

The right choice depends on what you are optimising for. Hermes is the best general-purpose answer if you want a persistent agent that lives across tools and platforms. n8n is the right answer if you want a flowchart of predictable workflows. The Claude Agent SDK is the right answer if you want pure reasoning quality and can write your own scheduling.

When to use Hermes Agent#

• You have a working SEO programme with real tools, real data and at least one senior person who understands the work.
• You want to automate the repetitive parts of that programme without giving up control.
• You can spend two weeks on initial setup and review the output weekly thereafter.
• You can budget £30 to £150 a month for model API and infrastructure.
• You will have human approval on anything that touches the live site.

When not to use Hermes Agent#

• You think it will write SEO content that ranks without editorial. It will not.
• You want to set it up once and never look at it. Agents drift without re-tuning.
• You do not have a senior SEO who can sanity-check what it produces. Without that layer, an agent on a live site is faster damage at lower cost.
• Your stack is small enough that the manual work takes 30 minutes a week. The setup will not pay back.

The pattern that works#

Hermes Agent is a powerful runtime in the hands of someone who already does SEO well. It is a way to give that person leverage over the repeated, error-prone, multi-tool jobs that fill an SEO week. Used that way, it earns its keep within a month.

Without that person in the loop, it is a very efficient way to systemically degrade a site's quality signals while showing you green dashboards. The framework will not protect you from yourself.

If you want to set this up with the SEO judgement baked in from the start, that is what I do for clients. Pricing is on the pricing page and the contact form is the way in.