dsh-fetch-third-party

When a model's search snippets are not enough to answer a question it needs to fetch the full page. Letting the local machine connect directly to arbitrary URLs, however, exposes an SSRF attack surface that adversarial prompts could exploit to target internal or loopback resources. dsh-fetch-third-party solves this with a courier model: target URLs are proxied exclusively through a user-configured third-party service (Tavily, Jina Reader, Firecrawl, or a custom contract endpoint), so the local runtime never initiates a direct outbound connection to the target, eliminating the SSRF risk by design.

The plugin is autonomous for the model yet fully controllable for the user. When snippets fall short the model invokes web_fetch_url on its own without user prompting. Multiple providers can be chained with ordered failover, exponential-backoff cooldown, and automatic recovery. API keys live only in a 0600 managed vault and are never echoed to the GUI, logs, or responses. Per-session budget caps, in-memory URL caching that skips both quota and budget, deep SSRF defense that rejects internal and reserved addresses before forwarding, structured output including title, headings, links, word count, and reading time, an optional local proxy, and automatic Crawl4AI container management with a live watchdog all round out the feature set.

It suits DeepSeek Harness users who want their model to read full articles rather than snippets, developers who refuse to lock fetching into a single SaaS, and self-hosters who prefer a zero-API-key stack. As long as you can point the plugin at any compatible fetch endpoint it becomes a ready-made deep-reading tool the model can reach for whenever it needs more context.