DonSeTch gives AI agents keyless web fetch, search and crawl in Rust

The developer dondai44423 has released DonSeTch, a Rust binary that gives any coding agent web fetch, search, and crawl through three tools, with no API keys and no third-party accounts required. The repository has gained 582 stars in its first month on GitHub.

What it does

DonSeTch exposes three tools over the Model Context Protocol: fetch a URL, run a search, and crawl a set of pages. It works with every MCP client the author lists, including Claude Code, Cursor, OpenCode, Pi, and Hermes, and it also runs as a standalone command-line tool. The tool definitions use about 3,500 tokens in total, so an agent that loads it does not lose much of its context window to them.

The fetch and crawl transport is written from scratch in Rust, using Chrome's own BoringSSL library directly rather than wrapping Playwright, Selenium, or the hyper HTTP library. The TLS ClientHello an agent sends is the one a real browser sends, so it does not become outdated the way a copied fingerprint table does when sites make their checks stricter. Session resumption, conditional revalidation, persistent cookies, and connection pooling are all present, which the author points out as the most obvious remaining signs of a bot, which most scraping tools skip.

Search without keys

The search tool queries ten or more backends in parallel and combines the results by cross-engine consensus and local semantic reranking, with no keys required for the default configuration. Optional bring-your-own-key adapters exist for teams that want to add a paid search provider.

Version 3 features

Version 3 focuses on saving agent context. Links render as short reference handles rather than full URLs, so a link that would cost about 80 tokens costs 3. A probe mode verifies whether a fully fetched page contains a claim and returns MATCH or NO-MATCH with up to three excerpts, in about 60 tokens instead of the four thousand a full page dump would use. A page-memory feature fingerprints every fetch and reports section-level diffs on re-fetches, which reduces a re-check to a single line when nothing has changed. Dead links can be resolved through the Wayback Machine, with the age of the snapshot printed alongside.

Install and license

DonSeTch installs from npm as a global binary that downloads a signed prebuilt for the current platform, or from Homebrew, or from source with Rust and Go toolchains. Linux ARM64 has no OCR or reranking because the ONNX prebuilt for that target deadlocks at load. The project is AGPL v3.

An agent that needs the live web usually needs three separate parts: a headless browser to get past bot blocking, a paid search API, and an extractor that turns a page into something the model can read. DonSeTch bundles all three into one binary and one MCP connection, and gives back handles and diffs instead of raw HTML so a long research session does not use up the context window. The trade-offs are AGPL licensing, which restricts what a proprietary product can include, and a new project that touches Chrome internals, which means compatibility with the next Chrome release is unproven. The 582 stars in thirty days show the problem it solves is real for the people who build agents that read.

Source

Primary source: dondai44423/donsetch on GitHub.