hermes-web-deepseek

DeepSeek Search + Fetch backend for Hermes Agent — routes the built-in web_search / web_extract tools through https://api.deepseek.com/v1/tools/{search,fetch}. No core changes, no wrapper skill.

What it gives you

search — DeepSeek's own index, up to 50 results per query, position-ordered
extract — server-side page fetch returning Markdown, up to 5 URLs concurrently
Per-URL failure isolation, one retry on 408/429/5xx, cancel-aware, honest supports_search() / supports_extract()

What DeepSeek opened up (2026-10-17 changelog)

From api-docs.deepseek.com/updates — "Opening up the in-house search API":

DeepSeek's web and API search capabilities are powered by our in-house DeepSeek search engine, which reliably handles hundreds of millions of search requests from around the world every day. DeepSeek Search is an AI-native general-purpose search engine that indexes hundreds of billions of web pages, supports multilingual search, and returns highly relevant, authoritative, and timely results in real time. It filters SEO spam, GEO manipulation, and low-quality AIGC through strict detection mechanisms.

That is the capability this plugin exposes to Hermes: the same engine behind DeepSeek's own web/APP search, reachable as web_search / web_extract.

Search queries forward their syntax intact, so the vendor's operators work unchanged: site:, -site:, "exact phrase", -term, lang:zh,en, after:2025-01-01, before:2026-06-01.

Results also carry rerank_score, authority (0–4), date and site_name; this provider maps summary → description and drops the rest, because Hermes' web_search row shape is title/url/description/position.

Install

hermes plugins install web-deepseek        # once the catalog entry is merged

Manual drop-in:

git clone https://github.com/Leegenux/hermes-web-deepseek \
  "$(hermes config env-path | xargs dirname)/plugins/web/deepseek"
hermes plugins enable web-deepseek

Put DEEPSEEK_API_KEY in .env (the same key the model API uses), then select it:

hermes config set web.search_backend deepseek-web
hermes config set web.extract_backend deepseek-web
# or both at once
hermes config set web.backend deepseek-web

The provider id is deepseek-web — not deepseek, which another catalog entry already registers (two providers cannot share an id; whichever loads last wins).

A gateway restart applies new plugin code; a fresh CLI session reads it automatically.

Cost and disclosure

$0.75 per 1,000 successful requests, per endpoint (Search and Fetch billed separately, per the official billing table). A request is billable only when it returns HTTP 200 and produces a result — failed requests, empty result sets and timeouts are free. count does not affect the price.
5 QPS per account per endpoint, search and fetch metered independently; Fetch additionally has a per-domain QPS limit, so a burst against one host can return 429 (the provider retries once). The provider runs at most 5 concurrent fetches, which is exactly the account budget.
Credentials: only DEEPSEEK_API_KEY, sent as a Bearer token to api.deepseek.com. Nothing is stored by the plugin.

Limitations (vendor-side, not fixable here)

Markdown only. fetch has no HTML/text variant, so web_extract(format="html") silently returns Markdown. The format kwarg is accepted and ignored on purpose.
Static HTML only. JavaScript-rendered pages and sites with anti-bot walls fail: 知乎 / X (Twitter) / nytimes.com all returned HTTP 404 in testing, while GitHub, arxiv, python docs, cloudflare blog and cls.cn worked.
Requests originate from DeepSeek's servers, not yours — so local DNS/proxy problems don't affect it, but sites that block DeepSeek's crawler will fail.

Development

Offline self-check (no network, no API key):

cd <this repo> && PYTHONPATH=<hermes install dir> python provider.py
# demo ok: search mapping, per-URL errors, order, timeouts, interrupt

Catalog admission check:

hermes plugins validate <this repo>

License

MIT — see LICENSE.

web-deepseek

What it adds

Environment variables it needs 1

README