Hermes plugin

Hermes Agent runs your model. tabstack-hermes is a Hermes plugin that hands the web work to Tabstack, so the model asks one question and receives a finished answer with its sources.

Hermes already ships web search and extraction backends. This plugin is not a replacement for search. It adds the calls that finish the job: a cited answer across sources, schema-enforced extraction, and multi-step browser tasks.

The plugin is open source under MIT at Mozilla-Ocho/tabstack-hermes and published to PyPI as tabstack-hermes.

---

Install

Requires Python 3.11 or newer, the same floor as hermes-agent, and an API key from the console.

pip install tabstack-hermes
hermes plugins enable tabstack
hermes env set TABSTACK_API_KEY <your-key>

Plugins are opt-in. The pip install puts the plugin on Hermes’ discovery path; hermes plugins enable tabstack is what lets it load.

To install from git instead of PyPI:

hermes plugins install Mozilla-Ocho/tabstack-hermes/tabstack_hermes --enable

The /tabstack_hermes suffix is the subdirectory holding the plugin. Hermes renames the installed directory to the manifest name, so it lands at ~/.hermes/plugins/tabstack/ either way.

Confirm it loaded:

hermes plugins list      # tabstack, enabled, 5 tools
hermes tools             # the tabstack toolset

The job: a current answer with sources

This is the one flow worth running first. Ask Hermes something its model cannot know, and watch which tool it reaches for.

> What changed in the most recent Node.js LTS release?

With the plugin enabled, the model calls research_question once. Tabstack plans the queries, finds and reads the sources, checks for gaps, iterates where it needs to, and returns a synthesized answer with the pages it cited. Your model receives the answer and its sources.

The tool returns a JSON string, which Hermes hands to the model:

{
  "answer": "Node.js 24 entered long-term support in October 2025...",
  "sources": [
    {
      "title": "Node.js Releases",
      "url": "https://nodejs.org/en/blog/release"
    }
  ]
}

What did not happen: no search-result ranking in the model’s context, no page fetching, no markup to recover readable text from, no second search to close a gap, and no citation assembly. The model spent its context on your conversation instead of on raw page content.

Without the plugin, the same question either runs through web_search and leaves the model to work the results, or gets answered from training data.

research_question takes an optional mode. The plugin defaults to balanced, which requires a paid plan; on Trial or Starter pass mode: "fast". Research bills per action, so a broader call costs more. See the Research guide.

The tools

All five land in the tabstack toolset, so they enable and disable as a unit in hermes tools.

Tool	What it does
research_question	Synthesized answer with cited sources across multiple pages.
extract_page_content	Fetch a page as clean markdown.
extract_structured_data	Pull specific fields from a URL into a JSON shape you define.
generate_structured_data	Fetch a page, then transform it into derived or reshaped JSON.
automate_browser_task	Run a multi-step, natural-language browser task on a public site.

Names, descriptions, and inputs match langchain-tabstack and the TypeScript adapters, so a Tabstack tool behaves the same whichever framework calls it. Tools return a JSON string; extract_page_content returns markdown directly.

Optional inputs

Sent only when the model provides them, so omitting them keeps Tabstack’s defaults. Exception: when mode is omitted, the plugin sends "balanced".

• extract_structured_data, extract_page_content, generate_structured_data:
  • effort: "min", "standard", or "max". Use "max" for JavaScript-heavy pages. See Effort levels.
  • nocache: true to bypass the cache.
  • country: ISO 3166-1 alpha-2 code for geotargeting.
• research_question: mode ("fast" or "balanced"), nocache.
• automate_browser_task: url (starting page), guardrails (constraints on what the agent may do), data (context for form filling), country, max_iterations, max_validation_attempts.

Tabstack as the web extract backend

The plugin also registers a tabstack web provider, so Hermes’ built-in web_extract tool can fetch through Tabstack without the model learning a new tool:

~/.hermes/config.yaml

web:
  extract_backend: "tabstack"

Extract only. Tabstack has no ranked search endpoint, so supports_search is False and web_search keeps using whichever backend you already have. For synthesis across sources, use research_question rather than a search backend.

How it behaves:

• URLs come back in the order they went in, because web_extract re-interleaves them with the ones it rejected as unsafe.
• A batch fans out 5 URLs at a time with a 60 second ceiling per URL. One failing URL returns an error entry for that URL and does not fail the batch.
• format="html" is ignored. Tabstack returns markdown.

Configuration

Variable	Purpose
TABSTACK_API_KEY	Required. Create one in the console.
TABSTACK_BASE_URL	Optional. Point the SDK at a different API base URL.

Keys are read through Hermes’ config layer first (~/.hermes/.env, written by hermes env set), then the process environment. Credentials therefore work in gateway sessions, delegated children, and subprocess agent runs where the variable was never exported.

Without a key the plugin still loads and the tools still appear in hermes tools, but a check_fn keeps them out of dispatch until a key is set. The SDK is imported and the client built on the first tool call, so a session that never calls Tabstack pays no cold-start cost.

Error handling

Handlers never raise. A failure returns JSON the model can act on, with the HTTP status when the API supplied one:

{ "error": "Extract failed for https://example.com", "status": 429 }

See the error reference for what each status means.

Limits

• automate_browser_task runs on public websites and cannot log in. For automation against a browser you control, see Pilo.
• The plugin covers extraction, not ranked search. Keep your existing web_search backend.
• Requests are hosted calls to the Tabstack API. See Data Handling for what is stored.

Next steps

• Research guide: modes, the event stream, and timeout strategy.
• Search versus research: which steps Tabstack runs inside the call.
• Other integrations: the same five tools in LangChain, Vercel AI SDK, Mastra, and more.