Hermes Agent becomes much more useful once it can do more than answer from the model’s existing knowledge. Connect it to a real search backend, give it a way to read full webpages, and add specialized tools for difficult sites, and it becomes a flexible research system.
That is what I build in this walkthrough.
The main setup uses:
- GPT-5.6 Sol or another capable model for reasoning
- SearXNG for free, multi-engine search discovery
- Tavily for reading and extracting selected webpages
- Hermes skills for links, briefs, and structured digests
- Bright Data MCP as an optional search and access fallback
I also test other models and tools to show that the system is not locked to one provider. Hermes can use different models for different jobs, select from available tools, and follow reusable skills that control how the final answer is presented.
The Basic Architecture
The most important part of this setup is understanding that search is not one action. It is a pipeline with separate layers.
1. The model reasons
GPT-5.6 Sol plans the research. It decides what to search, which sources look useful, when a page needs to be opened, and how the evidence should be organized.
The model is not the search engine. It is the reasoning layer controlling the process.
2. SearXNG finds sources
SearXNG discovers pages across multiple search engines. It returns titles, URLs, snippets, and sometimes the engines that found each result.
This is the discovery layer. It is fast and useful for finding current sources, but a search snippet is not the same as reading the complete page.
3. Tavily reads selected pages
Tavily opens selected URLs and returns cleaned page content. This gives Hermes more than a short snippet. The model can work from the actual article, documentation page, announcement, or product page.
This is the reading layer.
4. Hermes produces the answer
Once the sources have been found and read, the model compares the evidence, removes duplicate stories, identifies uncertainty, and produces the final response with links.
The result is a system where each layer can be changed independently. You can keep SearXNG and switch models. You can keep your model and use a different extractor. You can add specialist tools without rebuilding the entire workflow.
Step 1: Connect OpenAI Codex to Hermes Agent
Start by opening the Hermes Desktop App and going to Settings. Under providers, connect your OpenAI or ChatGPT account through the Codex login flow.
If you are new to the interface, start with my Hermes Desktop App walkthrough. It covers the app itself before you begin adding search tools.
After the account is connected, start a new Hermes session and choose the model you want to use.
For a broader look at the provider and its tools, read my full Codex in Hermes Agent walkthrough. I also have a separate guide showing how to use Codex in Hermes to create a marketing plan.
I use GPT-5.6 Sol in the walkthrough because it is a strong reasoning model, but it is not the only option. GPT-5.6 Terra or Luna can handle many search tasks. Lower-cost or faster models can also work well when the job is mainly tool calling and formatting.
The model’s job is to:
- Plan the search
- Choose tools
- Evaluate sources
- Decide when snippets are not enough
- Combine the results into a useful answer
You do not always need the largest model. A quick links search can run on a faster model. A larger model becomes more useful when the task requires comparison, synthesis, or deeper judgment.
Hermes Agent documentation: Read the official Hermes Agent documentation
Hermes Agent GitHub repository: View Hermes Agent on GitHub
Step 2: Install SearXNG With Docker
SearXNG is the free search workhorse in this setup. It is open source and can run locally in a Docker or Podman container.
The official SearXNG documentation recommends a containerized deployment for most users. A local manual instance can be exposed at:
You need Docker or Podman installed before running the container. The SearXNG configuration also needs JSON output enabled so Hermes can receive structured search results.
Instead of manually copying a wall of commands, give Hermes the official SearXNG repository and a clear outcome.
Use this prompt:
Install SearXNG from this GitHub repository and connect it to Hermes Agent. Use Docker, enable JSON search results, run it locally at http://localhost:8888, set it as the Hermes search backend, and verify it with a live search.
Official SearXNG repository: View the official SearXNG repository
Official container documentation: Read the official SearXNG container installation guide
Hermes can inspect the official project, identify the setup requirements, configure the local instance, and test the connection. If SearXNG is already installed, Hermes may detect the existing instance and verify it instead of reinstalling everything.
Once it is working, SearXNG gives Hermes a discovery layer without charging for every search query.
What SearXNG is good at
- General web discovery
- Current news searches
- Finding official documentation
- Returning raw links quickly
- Showing multiple sources for the same subject
- High-volume searches without a paid search API for every query
What SearXNG does not replace
SearXNG does not normally read every complete webpage it finds. It returns discovery data, including a title, URL, and snippet. For research that depends on the full article, Hermes needs a reading or extraction tool.
That is where Tavily comes in.
Step 3: Add Tavily for Full-Page Reading
Tavily gives Hermes access to the content inside selected webpages. It can search, extract, and return information in a format designed for AI applications.
Create a Tavily account, generate an API key, and add that key through the Hermes tools and API keys settings.
Do not paste a live API key into a public chat, screenshot, tutorial, or recording. Add it through the secure key interface and keep the value hidden.
Tavily dashboard: Open the Tavily dashboard
Tavily quickstart: Read the Tavily quickstart
After adding the key, use this prompt:
Use Tavily as my web extraction backend. Save the configuration and verify it by extracting the full content from a current news article.
The difference is simple:
- SearXNG finds the page
- Tavily reads the page
- The model analyzes the page
This matters when the details cannot be confirmed from a short search snippet. Product announcements, documentation, pricing pages, news reports, and competitor websites usually require full-page reading.
Step 4: Create Different Search Skills
Tools perform actions. Skills tell Hermes how to use those tools and how to present the result.
I use several SearXNG skills because I do not always want the same type of answer.
SearXNG Links
Links mode returns the raw discovery results:
- Title
- URL
- Search snippet
- Engine source when available
This is the fastest option when I want to review the search results myself. It avoids extra extraction work and keeps the response simple.
A request can be as direct as:
Search for the latest World Cup news and show me only the strongest links, titles, snippets, and engine sources.
SearXNG Brief
Brief mode turns the strongest results into a short conversational overview with source links.
This works well when I want more context than a raw list but do not need a full report.
Example:
Pull up a short brief on the latest World Cup news with links to the strongest sources.
SearXNG Digest
Digest mode produces a structured roundup. Each item can include a headline, source, short summary, and URL.
This is useful for recurring news reports, client updates, monitoring workflows, and content that needs to be shared with someone else.
For a working example of this type of recurring research, see my guide to setting up daily news brief cron jobs with Hermes Agent.
If the skills are not installed, ask Hermes:
Set up SearXNG search skills for raw links, short briefs, and structured digests. Verify that the skills are available and explain the trigger phrase for each one.
Skills are optional. If your default search and extraction backends are already configured, Hermes can still search from a normal request. Skills become valuable when you want consistent routing or a repeatable output format.
Step 5: Run a Live Search
The walkthrough uses current World Cup news because it makes the differences between links, briefs, extraction, and deeper analysis easy to see.
A complete research prompt looks like this:
Search for the latest 2026 World Cup news. First show me the strongest results with titles, URLs, snippets, and engine sources. Then open the three most important stories, verify the details from the full pages, and give me a short sourced brief.
This one prompt demonstrates the entire pipeline:
- The model plans the research.
- SearXNG discovers current stories.
- Links mode exposes the raw results.
- Tavily opens the strongest pages.
- The model compares the evidence and writes the final brief.
The final answer should include current publication dates, clickable source URLs, details verified from full pages, and a clear separation between confirmed facts and uncertainty.
Comparing Models for Agentic Search
The video also tests several models to show that search quality is not only about using the biggest option.
GPT-5.6 Sol
Sol works well as the main reasoning layer when the research requires planning, source comparison, and a polished final answer.
GPT-5.6 Luna and Terra
Luna and Terra can be good options for faster searches, simpler synthesis, and workflows where the tools are doing most of the retrieval work.
DeepSeek Flash
A fast DeepSeek model can be useful for straightforward discovery and tool-based research. In the walkthrough, Hermes uses the tools available in the session without requiring a specific search skill to be selected first.
GLM-5 Turbo
GLM-5 Turbo is tested with Tavily for a deeper World Cup request. The model calls the available tools, opens multiple articles, returns links, and produces a more detailed breakdown from the extracted pages.
The practical lesson is to choose the model based on the job. Use faster models for simple retrieval. Use stronger reasoning models when the output requires judgment, reconciliation, or strategic analysis.
If you want a wider model comparison, read GPT vs. Claude vs. Kimi vs. DeepSeek for business.
Adding Bright Data MCP as a Fallback
Bright Data is useful when a normal search or extraction path is not enough. Its MCP connection gives Hermes access to search, scraping, and browser-oriented tools.
In the walkthrough, Hermes identifies Bright Data MCP as an available tool and uses it without a manually selected skill. This demonstrates one of Hermes Agent’s strongest features: it can inspect the tools available in the current session and choose an appropriate route.
Bright Data is useful for:
- Search engine results
- JavaScript-heavy websites
- Pages that block ordinary extraction
- Geo-targeted results
- Browser-based access
- Batch search and extraction
Bright Data: Explore Bright Data
It does not need to replace SearXNG or Tavily. Treat it as another tool in the toolbox, especially when the normal path fails.
Other Search and Extraction Tools
The video mentions several additional options. Each one solves a different problem.
Exa
Exa focuses on semantic discovery. It can find companies, websites, and pages based on meaning rather than only exact keyword matches.
Best uses include competitor discovery, similar-company research, and finding sites within a category.
Firecrawl
Firecrawl becomes useful when the job requires crawling multiple pages from one website instead of reading one selected URL.
Best uses include documentation sites, competitor website analysis, content inventories, and clean Markdown extraction.
Brave Search and DuckDuckGo
These can provide additional search routes or fallbacks. They are useful when you want different coverage from a standard discovery query or prefer not to rely on one search source.
The goal is not to install every provider. Start with one working pipeline and add specialized tools when the research demands them.
Recommended Starting Stack
For most users, I would start with:
- Hermes Agent or the Hermes Desktop App
- A connected model provider
- SearXNG running locally through Docker
- Tavily for full-page extraction
- One links skill and one brief skill
- Bright Data MCP only when you need a stronger fallback
That setup covers general discovery, current news, source verification, full-page reading, and structured answers without turning the configuration into a catalog of every available tool.
Final Takeaway
Hermes Agent becomes a search powerhouse when you stop treating search as one feature and start treating it as a system.
The model reasons. SearXNG finds. Tavily reads. Bright Data unlocks harder sources. Skills control the workflow and output.
You can use GPT-5.6 Sol, a faster GPT model, DeepSeek, GLM, or another provider. You can ask for raw links, a short brief, or a structured digest. You can let Hermes choose tools automatically or tell it exactly which route to use.
That flexibility is the real advantage. It is not one model connected to one search engine. It is a research toolbox you can shape around the work you actually do.
If you want help building AI research, content, SEO, or marketing automation workflows for your business or agency, review my digital marketing and AI automation services or contact Jason Pollak Marketing.
Agencies can also review my white label marketing services for behind-the-scenes AI, SEO, content, and marketing execution.
Schedule a strategy call: Book a free 30-minute strategy call
Related Guides From Jason Pollak Marketing
- How to Use Codex in Hermes Agent
- How to Use Codex in Hermes Agent to Create a Marketing Plan
- Hermes Desktop App: The Open-Source AI Agent That No Longer Needs a Terminal
- How to Set Up Daily News Brief Cron Jobs With Hermes Agent
- Why AI Agents Like Hermes Are the Future
- How to Run a 21-Job AI Automation Pipeline for Pennies a Day