Skip to content

New: deep research, balance and auto-reload tools in the keyless MCP server. Install in one line

Changelog

Product updates to the ContextFetch API, MCP server and site. Numbers come from DONE.md and the benchmark files.

  1. Docs, playground and blog

    The docs moved from one long page to a sidebar of focused pages, all generated from the same API reference file the llms.txt files use.

    • Playground: a full-page keyless playground for /v1/search, /v1/contents and /r/, with latency and token chips.
    • Blog: three engineering posts written from our own measurements.
    • Old links: anchors such as /docs#payment now forward to their new page.
    Open the playground →

  2. Live at contextfetch.dev

    The site and the API went to production, with live Stripe checkout behind the buy links.

    • API: api.contextfetch.dev on a dedicated server (Ryzen 7 3700X, 64 GB RAM) behind Caddy with Let’s Encrypt.
    • Checkout: /buy/{instant,builder,scale} answer 303 straight to Stripe Checkout.
    • First live run: five queries from a client in Florida all returned 10 results; cold searches took 2.2 to 3.5 s and cached repeats 143 to 282 ms, network included. The index held 40,316 documents.
    Read about the latency work →

  3. Renamed to ContextFetch

    The product took its final name and domains. Every user-visible name changed in one pass.

    • MCP package @contextfetch/mcp with 6 tools: contextfetch_search, contextfetch_extract_url, contextfetch_find_similar, contextfetch_deep_research, contextfetch_balance, contextfetch_auto_reload.
    • The response extension field is now contextfetch, same shape as before.
    • Fetches identify as ContextFetchBot/1.0 (+https://contextfetch.dev/bot).

  4. Packs, auto-reload and the 402 loop

    Prepaid packs replaced the earlier two-tier pricing, and agents got everything they need to pay without a browser session of their own.

    • Packs: Instant $15 for 15,000 credits, Builder $50 for 60,000 credits, Scale $150 for 200,000 credits. One-time payments; credits never expire.
    • Keyless: 25 requests per IP per UTC day, X-Credits-Remaining on every billable response, and a 402 with X-Payment-URL, the packs and a one-click payment link when it runs out.
    • Auto-reload: opt-in, with a threshold and a monthly cap. GET /v1/balance and POST /v1/billing/auto-reload, also as MCP tools.
    Read about the 402 loop →

  5. Faster cold search, more sources, two caches

    Cold search p50 on our M2 test machine went from 4,140 ms to 1,644 ms (max 6,356 to 2,300 ms) over five fixed queries. We missed our 1.2 s target and say why in the post.

    • Recall: web, Wikipedia, Hacker News, Stack Exchange, GitHub, arXiv and the local index are queried at once under a time budget and fused with reciprocal rank fusion.
    • Fetch and rerank: pipelined fetching with a 950 ms phase budget, then one batched cross-encoder call per request, capped at 256 tokens per pair.
    • Caches: a 15-minute query cache and a 48-hour URL content cache. A cache hit takes 1 to 4 ms at the server.
    • Response field: cache state, stage timings, token counts and per-source hits on every search. MCP: a deep research tool that runs up to 3 searches in parallel.
    • Pruning: on average 69% fewer tokens than the full page (median 79%) in a benchmark on 29 synthetic docs; not yet measured on live pages.
    Response field docs →

  6. The API

    The first complete version of the engine, the site and the MCP server.

    • Exa-compatible /search, /contents and /findSimilar, Tavily-compatible /tavily/search and the /r/ reader. Real exa_py and tavily-python calls parsed without errors in our full-stack test.
    • Highlights and a cross-encoder-pruned context on every result, in the same credit as the search.
    • Safety: an SSRF guard on every user-supplied URL, keys stored as SHA-256 hashes, refunds when a paid request fails, and keyless limits per IPv6 /64.
    How we benchmark →