osaurus-ai/osaurus-tools/fetch

Fetch

Lightweight HTTP client for grounding agent work in real web content. Hardened by default: SSRF guard blocks private IPs, response size capped at 10 MB, downloads sandboxed to ~/Downloads.

When to use

• Calling a public JSON API (fetch_json).
• Reading the content of a static page or article (fetch_html returns Markdown).
• Saving a file to disk (download).
• Anything where the page does not require JavaScript execution or a real browser session.

When NOT to use

• Pages that need login, JavaScript rendering, or interaction → use osaurus.browser.
• Web search → use the host's built-in web_search (then feed top URLs to fetch_html).
• Reaching localhost / 127.0.0.1 / private LAN IPs → blocked by SSRF guard. Set allow_private: true only if the user explicitly asked for it.

Canonical workflow

1. web_search(query)                      → list of URLs
2. osaurus.fetch.fetch_html(url)          → { markdown, title, byline, ... }
3. Reason over the markdown

Or for APIs:

1. osaurus.fetch.fetch_json(url, headers, body)
2. data is already parsed JSON in the response

Output envelope

{
  "ok": true,
  "data": {
    "status": 200,
    "final_url": "https://example.com/after/redirects",
    "redirect_chain": ["https://t.co/abc", "https://example.com/..."],
    "protocol_version": "h2",
    "headers": { "content-type": "..." },
    "body": "...",                     // fetch only
    "json": { ... },                   // fetch_json only
    "markdown": "...", "title": "...", // fetch_html only
    "truncated": false
  }
}
{ "ok": false, "error": { "code": "...", "message": "...", "hint": "..." } }

Common error codes: INVALID_ARGS, SSRF_BLOCKED, TIMEOUT, RESPONSE_TOO_LARGE, HTTP_ERROR, EXTRACTION_FAILED, DOWNLOAD_PATH_INVALID.

Tips

• For long pages, prefer fetch_html over fetch — the Readability extractor returns clean Markdown a model can actually use, while fetch returns the raw HTML body.
• auth: { type: "bearer", token: "..." } is a shortcut for an Authorization header — use it instead of constructing headers manually.
• download rejects path separators, .., and absolute paths in filename. Pass a plain filename only.
• max_bytes caps the in-memory response. Default 10 MB. If you hit truncated: true, your data is incomplete.