If your DSH can't read the web, this plugin is for you: Stealthy fetch, no API keys, free, easy setup

If your using dsh or any other coding harness, i recommend trying out donsetch, it will easily solve your agent's web research problem in like 2 minutes, its dead simple as well, Dosent require any API, Works out of the box. As you can see the demo above, it is using local keyless search, i removed my byok to demonstrate local keyless search is all you need if you are not doing super heavy research. Native dsh plugin, not a generic MCP config: dsh plugin --profile web add github:dondai44423/donsetch-dsh Downloads the prebuilt binary, verifies the sha256 Registers donsetch_fetch , donsetch_search , donsetch_crawl as native tools Auto-updates with every release Restart dsh web once. Your agent can read the web. Why the fetch tier matters Most tools fake a browser fingerprint with tables that rot. DonSeTch drives Chrome's actual BoringSSL, with its own HTTP/2 stack on top. The h2 preface is CI-asserted byte-identical to real Chromium. When it hits a bot wall, a headless browser wakes up, solves the challenge, hands the clearance cookies back to the fast HTTP path, and goes back to sleep. The browser almost never fetches content. Benchmark: 48 URLs across three anti-bot tiers. 95.8% retrieval, median 772ms, zero false positives. It reports blocked honestly, it never wraps a captcha page in a success blob. Search, keyless Bing, DDG, Brave, Mojeek, Yahoo in parallel, plus GitHub, HN, arXiv, StackExchange, MDN verticals. Consensus ranking + local ONNX reranker. Free forever. A URL returned by three independent index families gets boosted Dead engines get benched and stay benched across restarts When egress comes back thin, a ghost lane re-renders Google's SERP through the browser 95.5% answer-in-snippet on a 110-question benchmark vs Tavily's 93.3% Have Exa/Tavily/Serper keys? Optional BYOK (can stack multiple keys as well from same provider, two 1500 credit accounts become 3000) Don't? Nothing breaks. Built for agents in the loop focus : ask "memory safety" on a 50KB page, get the relevant section. 80% context cut Handles : search returns S1...S12 , then fetch S3 just works Probe mode : verify a claim for ~60 tokens instead of 4k Dead links auto-resurrect from Wayback PDFs, scanned included : pixel-fusion extraction that can't hallucinate text Login sessions : donsetch login x.com , you sign in normally, your credentials never enter the tool Limits, stated plainly Interactive captchas are a deliberate dead end If you dont configure Bright data Web solver(clean error, never a hang). Not for bulk scraping. AGPL-3.0 (if you care about that) Repo (MCP, CLI, Homebrew, DSH plug in, Pi extension): github.com/dondai44423/donsetch Star the the project if you like it 😄