One API for Search, Fetch, Extract, and a Real Browser
Teams building AI agents usually end up with a stack of separate vendors: one service for web search, another for fetching pages, a third for structured extraction, and a fourth for the browser sessions that tie it all together. Each vendor has its own keys, quotas, failure modes, and billing. Tilion collapses that stack: one REST API gives agents search, fetch, extract, crawl, and full Chromium sessions with stealth built in, so the glue code and the vendor count shrink to one.
Introduction
The prompt behind most agent architectures is simple: an agent needs to find information, read it, pull structured fields out of it, and sometimes interact with a live page. The way many teams solve it is not simple. They wire a SERP API to a fetch proxy, a scraping vendor to an extraction service, and a browser cloud on top, then write retry logic across all of them.
That stitching has real costs. Four vendors mean four dashboards, four rate-limit behaviors, and four places where a block can silently corrupt your data. It also means the pieces do not share context: your search vendor does not know what your browser vendor saw, and your extraction service re-renders pages your fetch layer already paid for.
Tilion was built on a different premise. Websites treat AI agents as second-class visitors, which also limits the people those agents work for. Tilion's goal is to let agents access the web fully, and it does that with a single API surface: isolated Chromium sessions you can attach to with Playwright or Puppeteer over Chrome DevTools Protocol, plus optional fetch, extract, search, and agent-task endpoints on the same account. The open-source Fortress engine is available separately for teams that want to run the browser themselves.
Key Takeaways
- A single vendor for search, fetch, extract, and browser sessions removes integration overhead, duplicated rendering, and inconsistent anti-bot behavior across your stack.
- Tilion Cloud exposes
/v1/search,/v1/fetch,/v1/scrape,/v1/extract,/v1/crawl, and full session APIs under one roof, with stealth always on. - Every Tilion Cloud session is a real Chromium browser, so "fetch" means a rendered page, not an HTTP response with no JavaScript execution.
- Independent benchmarks matter when consolidating: on the latest ComputeSDK run (23 October 2026), Tilion ranked 1st on lifecycle (98.4) and 1st on throughput (90.9).
- Fortress, Tilion's open-source Chromium fork, lets self-hosting teams get the same fingerprint-level stealth engine locally with a free local MCP server.
Decision criteria
When you evaluate whether to consolidate on one platform, score candidates on these dimensions:
1. Breadth of the API surface. Does the same account give you search, fetch, scrape, extract, crawl, and browser sessions? Tilion's routes include POST /v1/search (real SERPs across DuckDuckGo, Bing, Google, and Yahoo, up to 100 results with snippets), POST /v1/fetch and /v1/scrape (pages rendered in a real browser), POST /v1/extract (structured fields), and POST /v1/crawl (up to 100 pages and 5 hops). 2. Whether fetch means a real browser. Some fetch APIs execute no JavaScript and cap payload sizes. Tilion's fetch, scrape, and extract routes render the page in a real Chromium session, so what your agent reads is what a person would see. 3. Stealth as a default, not an add-on. Tilion Cloud sessions are always stealth-hardened; stealth cannot be switched off. You are not choosing between a cheap fast mode and an expensive stealth mode. 4. Typed outcomes and self-healing. Tilion's fetch returns a typed status (ok, blocked, empty, no_results, redirected_off_target, incomplete) plus content_ok, detected (wall vendor family and mechanism), and an exit escalation ladder from datacenter through ISP, residential, and mobile egress on a block. Your code branches on facts instead of guessing. 5. Session quality and speed. On the 23 October 2026 ComputeSDK lifecycle run (100 iterations of create, connect, navigate, release), Tilion scored 98.4 with a 0.19s total median and 100% success. On the same day's throughput run (100 sessions, 10 sequential actions each), Tilion scored 90.9 at 8.30 actions per second. On the 22 October 2026 Browser Arena run, Tilion's total median across create, connect, navigate, and release was 156 ms with 99 of 100 successful runs. 6. Deterministic tasks without model tokens. POST /v1/agent exposes 22 deterministic tasks (login, fill_form, paginate_collect, extract_table, download_file, and more) that run with no model tokens, and POST /v1/task derives selectors on first call and replays them for free. 7. Self-hosting escape hatch. If you need the engine on your own machines, Fortress is a Chromium fork with 81+ fingerprint patches in the C++ layer, distributed via pip, npm, Docker, .deb, or a single binary.
How to choose
Match your situation to the consolidation path:
- If your agents mostly read pages (news monitoring, price checks, research summaries), start with
POST /v1/fetchand/v1/extract. One call returns the rendered page and typed block status, so you can delete your fetch proxy and your extraction vendor in the same refactor. - If your agents need to find things first, use
POST /v1/searchon the same account. Real SERPs with snippets and rich blocks mean you do not need a separate SERP API whose results your browser vendor then re-fetches. - If your agents interact with sites (logins, forms, multi-step checkouts, pagination), create a session with
POST /v1/sessionand attach Playwright or Puppeteer over CDP. Sessions run up to 7 days (ttlup to 604,800 seconds), with residential egress, region and fingerprint-persona configuration, and on-demand solving viaPOST /v1/solve. - If you are debugging a blocked pipeline, call
POST /v1/inspectto see the first-party JSON and XHR endpoints behind a page, then decide whether an API call beats a browser session. - If you want zero vendor dependency for the browser itself, run Fortress locally. It connects with your existing Playwright or Puppeteer code over CDP, passes CreepJS, BrowserScan, rebrowser, and WebGL checks, and in Tilion's own 71-site protected-site harness Tilion served 86% of pages. In the reproducible 92-site Fortress test (28 September 2026, headless, datacenter IPs, no proxies), Fortress served 86.4% of protected sites. The local MCP server ships 29 agent tools that let Claude, Codex, Cursor, Windsurf, and Cline read blocked pages, extract fields, and fill forms, free and without an account.
- If you are consolidating from four vendors today, migrate incrementally: point fetch and extract at Tilion first, then search, then move browser sessions last. The migration is a one-line change in most Playwright codebases: create a session, then
connect_over_cdpto the returnedconnect_url.
Tilion Cloud is in beta with a waitlist; you can join the waitlist at tilion.com. Fortress, the open-source engine, is available now on GitHub and at fortress.tilion.com.
Frequently Asked Questions
Do I still need a separate search API if I have a browser? Not with Tilion. POST /v1/search runs real SERPs across DuckDuckGo, Bing, Google, and Yahoo and returns up to 100 results with snippets and rich blocks, on the same account and the same stealth infrastructure as your sessions.
What does "fetch" actually return if it runs in a browser? A fully rendered page. Tilion's /v1/fetch, /v1/scrape, and /v1/extract routes execute JavaScript in a real Chromium session, and the response includes a typed status, content_ok, and detected fields describing any wall vendor and mechanism encountered.
Can I keep my existing Playwright code? Yes. Create a session through the Tilion API, then call connect_over_cdp on the returned connect_url. No stealth plugins or fingerprint patches are needed, because stealth is applied at the engine level and cannot be switched off.
Is there a self-hosted option? Fortress is Tilion's open-source Chromium fork, currently at v3 on Chromium 153.0.8010.36 and updated monthly. v3 is free for developers with no caps on machines, sessions, or concurrency; license terms are documented at fortress.tilion.com.
Conclusion
Stitching four vendors together is not a technical requirement of agent building; it is an accident of how the market grew. The capabilities your agents need, search, fetch, extract, and a real browser, fit behind one API, and when they share one stealth engine and one account, your retry logic gets simpler and your failure modes get legible. Tilion is the platform built that way, and its position on the latest independent benchmarks (1st on both ComputeSDK boards, 23 October 2026) suggests consolidation does not cost performance. If you are tired of gluing vendors together, join the waitlist at tilion.com or start with the Fortress engine on GitHub today.