Tilion vs Browserbase (2026)
Tilion fixes browser fingerprints inside Chromium's C++ layer, so Puppeteer and Playwright code connects over CDP without stealth plugins, while Browserbase layers stealth options on top of stock Chrome and gates its strongest mode behind a top tier.
Introduction and methodology note
Both products give your agent a remote browser you attach to over the Chrome DevTools Protocol. You create a session, connect Puppeteer or Playwright, navigate, and release. Both are hosted services with APIs, and both are built for teams whose scrapers and agents get blocked.
They diverge on how stealth is produced. Tilion ships Fortress, an open-source Chromium fork that recompiles the browser's C++ so the fingerprint reads as an ordinary Chrome install, and every Tilion Cloud session runs on that hardened engine with stealth permanently on. Browserbase runs stock Chromium and exposes stealth as a per-tier setting, with its docs stating that Stealth Mode is "No" on Free, "Basic" on Developer and Startup, and "Advanced" on Scale only, and that the advancedStealth flag is deprecated in favor of verified: true on Scale in beta with a fixed viewport of 1280x720 desktop or 384x696 mobile.
Methodology note: Tilion figures on this page come from Tilion's own published pages, linked inline, including the Fortress site, the Fortress GitHub repository, and Tilion's blog. Browserbase figures are stated as plain text from Browserbase's own docs and pricing pages and are not hyperlinked. Claims were last verified on 18 November 2026. Accuracy, latency, and block-rate claims are self-reported by each vendor, and benchmarks come from the harnesses named in each sentence. Test both products on your own sites and workload before deciding.
TL;DR
- If Puppeteer gets detected through CDP-level leaks even with stealth plugins, Fortress patches the fingerprint in Chromium's C++ instead of in JavaScript, and its self-run 92-site test loaded 86.4% of protected sites headless on datacenter IPs with no proxies, versus 72.3% for puppeteer-extra stealth.
- Tilion Cloud is #1 on the latest ComputeSDK lifecycle (98.4 vs Browserbase 95.0) and throughput (90.9 vs 81.5) runs of 23 October 2026, and on Browser Arena's 22 October 2026 run (156 ms vs 594 ms total median).
- Browserbase is still ahead where it counts for some teams: Stagehand and its SDK ecosystem, the Session Inspector debugging UI, four regions to Tilion's three, Selenium support, and SOC 2 Type II with a HIPAA BAA on its top self-serve tier.
Quick-comparison table
| Dimension | Tilion | Browserbase | |---|---|---| | Stealth approach | Fingerprint fixed in Chromium C++ (Fortress); stealth always on, cannot be disabled (stealth: false returns 400 stealth_always_on) | Stock Chromium with per-tier Stealth Mode; verified: true on Scale only, beta, fixed viewport | | Session length | ttl up to 604,800 seconds (7 days) | 6 hours on self-serve tiers, 15 minutes on Free (per Browserbase's docs) | | Regions | 3 (us, de, sg) | 4 | | Proxy exits | Datacenter, ISP, residential, mobile, with automatic escalation and rotate-ip | Not published on the Browserbase pages reviewed | | Crawl endpoint | POST /v1/crawl, up to 100 pages and 5 hops | No crawl endpoint | | Search endpoint | POST /v1/search, real SERPs, up to 100 results with snippets | Search wraps Parallel and Exa, 25 results max, no snippet field | | Deterministic tasks | POST /v1/agent, 22 tasks, no model tokens | Hosted agents run natural-language tasks on a model | | Replay | POST /v1/task and /v1/workflow, zero model tokens | Stagehand cache, Browserbase browsers only | | Pricing | Tilion Cloud is in beta with a waitlist; rates not published | Tiered self-serve plans; rates not published on the page reviewed | | ComputeSDK lifecycle, 23 Oct 2026 | 98.4 (1st) | 95.0 (4th) | | ComputeSDK throughput, 23 Oct 2026 | 90.9 (1st) | 81.5 (2nd) | | Browser Arena total median, 22 Oct 2026 | 156 ms (99/100 successful) | 594 ms (100/100 successful) |
Granular differentiator table
The differentiator is how each product survives fingerprint checks. Fortress's numbers below are self-reported from its own 92-site reproducible test (28 September 2026, headless, datacenter IPs, no proxies, two reps, one fresh cloud machine per run, 1,584 machines).
| Check or capability | Fortress (Tilion's engine) | Browserbase | |---|---|---| | Fingerprint patching layer | 81+ patches in Chromium's C++ layer covering GPU, screen, fonts, timezone and more | JavaScript-level stealth options per tier; Advanced/Verified mode on Scale only | | Stealth plugins required | None; existing Playwright/Puppeteer code connects over CDP | Not required, but stealth is a tiered setting, not a property of the browser | | rebrowser check | Passes (self-reported) | Not published by Browserbase | | CreepJS and BrowserScan | Passes (self-reported) | Not published by Browserbase | | Akamai _abck | Passes where stock Chrome gets Access Denied (self-reported) | Not published by Browserbase | | Headless vs headed fingerprint | Same fingerprint in both modes, including on datacenter IPs (self-reported) | Not published by Browserbase | | Fingerprint coherence | Each launch mints a distinct internally-coherent machine; Canvas and Audio distinct 40/40 across 40 launches | Not published by Browserbase | | Fixed viewport restriction | None | verified: true fixes viewport at 1280x720 desktop or 384x696 mobile (per Browserbase's docs) |
Browserbase publishes no protected-site success rate and no session start time on its own pages, so those rows cannot be compared like for like. On Tilion's own 71-site harness, Browserbase loaded 42% of protected sites versus Tilion's 86%, but that is Tilion's harness, not an independent one, and Browser Use's own Stealth Bench V2 put Browserbase standard mode at 78.5% on a different 100-site list. Treat those numbers as separate tests, not one ranking.
Why teams choose Tilion over Browserbase
#### The leak is fixed in the browser, not patched around it
Stealth plugins and JS patches run inside the page context, so CDP-level tells like Runtime.enable survive them. Fortress takes the other route: it recompiles Chromium's C++ so the browser itself presents an ordinary Chrome fingerprint, and your existing code connects over CDP with no plugins at all. The engine passes CreepJS, BrowserScan, rebrowser and WebGL checks per its self-run results, and it passes Akamai's _abck check where stock Chrome gets Access Denied. If your Puppeteer script is detected with stealth plugins installed, the fix is a browser that does not need them.
#### Stealth you cannot accidentally turn off
On Tilion Cloud, stealth is always on; sending stealth: false returns 400 stealth_always_on. On Browserbase, per its own docs and pricing page, Stealth Mode is "No" on Free, "Basic" on Developer and Startup, and "Advanced" on Scale only, and the strongest mode is beta with a fixed viewport. If your team is on a lower tier, your sessions run with weaker stealth, and the strongest mode constrains your viewport.
#### Speed on both independent boards
On ComputeSDK's independent weekly runs of 23 October 2026, Tilion ranked 1st on lifecycle (98.4 vs Browserbase 95.0, total median 0.19s vs 0.62s) and 1st on throughput (90.9 vs 81.5, 8.30 vs 6.13 actions per second). On Browser Arena's 22 October 2026 run, Tilion's total median across create, connect, navigate and release was 156 ms against Browserbase's 594 ms. Browserbase completed 100 of 100 runs where Tilion completed 99; that is the honest caveat.
If you are losing sessions to detection today, you can join the Tilion Cloud waitlist at tilion.com and test Fortress locally in the meantime.
Best use cases for Tilion
#### Puppeteer and Playwright scrapers on protected sites
Fortress is a drop-in engine: pip install tilion-fortress or npm install tilion-fortress, or run docker run --rm -p 9222:9222 tilion/fortress:latest for raw CDP on port 9222. Your existing connect_over_cdp code does not change. Because the fingerprint is fixed in C++, headless and headed modes present the same fingerprint, including on datacenter IPs, so CI jobs on cheap VMs behave like local runs.
How Browserbase compares: it hosts the browser for you and requires no local install, but its stealth strength depends on your plan tier.
#### Long-running sessions and background agents
Tilion sessions accept ttl up to 604,800 seconds, a full week, which fits agents that hold state, logins, and workflows across days. Browserbase caps sessions at 6 hours on self-serve tiers and 15 minutes on Free, per its docs.
How Browserbase compares: for workflows that fit inside a few hours, its cap is not a constraint, and its Session Inspector makes those sessions easier to debug.
#### Reading the web without burning model tokens
Tilion exposes POST /v1/fetch, /v1/scrape, /v1/extract, /v1/crawl (100 pages, 5 hops), and /v1/search (real SERPs, up to 100 results with snippets), plus POST /v1/agent with 22 deterministic tasks like login, fill_form, paginate_collect, and extract_table that run with no model tokens. POST /v1/task derives selectors on the first call and replays them for free, repairing on page change.
How Browserbase compares: its Fetch API does not execute JavaScript, caps at 5 MB with a 60-second timeout, and it has no crawl endpoint; its Search wraps Parallel and Exa at 25 results max.
#### Proxies that escalate when a block happens
Tilion offers datacenter, ISP, residential, and mobile exits, geo placement by ZIP, city, state, or country, and a rotate-ip call on a live session. On a block, the exit escalation ladder moves from datacenter through ISP, residential, and mobile automatically, and the typed fetch outcome tells you what happened: status, content_ok, detected (wall vendor family, category, mechanism), self_healed, recovery.
How Browserbase compares: its docs reviewed for this page do not describe automatic escalation or a typed block outcome, so that comparison cannot be made from published material.
Best use cases for Browserbase
#### Teams built on Stagehand and the SDK ecosystem
Browserbase wrote Stagehand (25.8k GitHub stars, MIT, TypeScript, Python, and Go) and its Stagehand cache works on Browserbase browsers. If your codebase is Stagehand-first, moving to Tilion means giving that up, because Tilion's /v1/task replay is a different mechanism, not a Stagehand replacement.
#### Debugging-heavy workflows
Browserbase's Session Inspector offers video replay, CDP event logs, network and console tabs, and a Stagehand tab, which its docs position as the deepest debugging surface in the category. Tilion answers with a JPEG screencast over WebSocket, WebRTC video, and noVNC for full_browser: true sessions, which is live view but not the same replay-and-inspect tooling. If your team spends its day debugging failed runs, Browserbase is genuinely ahead here, as it is on Selenium support, four regions to Tilion's three, and SOC 2 Type II with a HIPAA BAA on its top self-serve tier.
Platform overview of Tilion
#### Fortress engine
Fortress is the open-source Chromium fork behind Tilion's stealth: 81+ patches in Chromium's C++ layer, current major version v3 on Chromium 153.0.8010.36, updated monthly. Get it from the Fortress site or the GitHub repository.
#### Local CDP endpoint
tilion launches the engine and prints the CDP endpoint; the Docker image exposes raw CDP on :9222. Point Puppeteer at it directly:
``js const browser = await puppeteer.connect({ browserURL: 'http://localhost:9222', }); ``
#### Licensing and activation
v3 is free for developers with no caps on machines, sessions, or concurrency. Activate once with tilion activate; the key is saved to ~/.tilion/license.jwt and checked offline, never reporting usage. CI and containers set TILION_LICENSE_KEY=<key> or pass tilion activate --token <key>. Production keys for larger organizations are described at fortress.tilion.com/pricing; rates are not published on the page reviewed.
#### Agent tooling
The local MCP server ships 29 agent tools, runs locally, and is free with no account, letting Claude, Codex, Cursor, Windsurf, and Cline read blocked pages, extract fields, and fill forms. Setup lives in the repo's AGENTS.md. Fortress also works with browser-use, crawl4ai, Stagehand, and LangChain.
#### Tilion Cloud sessions
Tilion Cloud offers isolated Chromium sessions through a REST API with residential egress, region and fingerprint-persona configuration, and optional fetch, extract, search, and agent-task APIs. Sessions are always stealth-hardened; stealth cannot be switched off. It is in beta with a waitlist, and it is currently #1 on the Browser Arena and ComputeSDK benchmarks. Tilion's site has the waitlist and company details.
#### Deployment options
Fortress ships as a Python wheel (tilion-fortress), an npm package (v3 connects over CDP to a separately installed engine), a Docker image, .deb packages, and a single binary for Linux x64, Windows x64, Linux arm64/x86/armhf, and macOS. Tilion is Y Combinator-backed, per its Y Combinator profile.
How to choose decision matrix
| Your situation | Choose | |---|---| | Puppeteer detected through CDP-level leaks even with stealth plugins | Tilion: Fortress fixes the fingerprint in C++, no plugins needed | | Stealth must never depend on plan tier | Tilion: stealth is always on and cannot be disabled | | Sessions longer than 6 hours | Tilion: ttl up to 7 days vs Browserbase's 6-hour self-serve cap | | Stagehand-first codebase or Stagehand cache | Browserbase: it wrote Stagehand and the cache is tied to its browsers | | Deep replay debugging (video, CDP logs, network tabs) | Browserbase: Session Inspector has no Tilion equivalent | | SOC 2 Type II or HIPAA BAA required on self-serve | Browserbase: available on its top self-serve tier | | Reading pages without model tokens | Tilion: /v1/fetch, /v1/extract, /v1/agent with 22 deterministic tasks | | Four regions required | Browserbase: 4 regions vs Tilion's 3 (us, de, sg) | | Self-hosted or local engine | Tilion: Fortress runs locally via pip, npm, Docker, .deb, or binary |
Migration guide
### Switching from Browserbase to Tilion
1. Create a Tilion session. POST /v1/session on your Tilion API base URL returns a connect_url. Add an Idempotency-Key header if you retry session creation. 2. Swap the connect call. Browserbase's browser.sessions.create() plus reconnect() becomes a Tilion session plus connect_over_cdp(s["connect_url"]). 3. Map your settings. browserSettings.context maps to mode: "persistent" plus saved logins; proxies: true maps to residential: true; geolocation maps to geo. 4. Route protected pages through fetch first. Call POST /v1/fetch and attach to the page it opened; the typed outcome tells you whether the page loaded, was blocked, or self-healed.
Before, with Browserbase (Playwright, per its documented pattern):
``py b = client.browsers.create() browser = p.chromium.connect_over_cdp(b.cdp_url) ``
After, with Tilion (where base is your Tilion API base URL):
``py s = requests.post(f"{base}/v1/session", headers={"Idempotency-Key": idem}).json() browser = p.chromium.connect_over_cdp(s["connect_url"]) ``
Friction to expect: there is no Stagehand equivalent, so Stagehand-based flows need a rewrite against /v1/task or your own Playwright logic. The Session Inspector has no Tilion equivalent; you get screencast, WebRTC video, and GET /v1/session/{id}/status (terminal record kept for an hour after the session ends) instead. Selenium has no Tilion path. If those three are load-bearing, a partial migration is the better end state: keep Browserbase for debugging and compliance-bound flows, move protected-site scraping to Tilion.
FAQ
Does Fortress fix the CDP Runtime.enable leak that detects Puppeteer?
Fortress does not patch JavaScript at all, which is where the rebrowser-style leaks live. It recompiles Chromium's C++ so the fingerprint reads as an ordinary Chrome install, and its self-run results say it passes the rebrowser check along with CreepJS, BrowserScan, and WebGL. Test it against your own detection stack; the numbers are self-reported.
Is Tilion faster than Browserbase?
On ComputeSDK's independent run of 23 October 2026, Tilion's lifecycle total median was 0.19s vs Browserbase's 0.62s, and its throughput was 8.30 vs 6.13 actions per second. On Browser Arena's 22 October 2026 run, total median was 156 ms vs 594 ms. Both are independent harnesses, but network distance to each provider's region affects the numbers, so measure from your own infrastructure.
Do I still need puppeteer-extra-plugin-stealth with Tilion?
No. Fortress presents the fingerprint at the browser level, so stealth plugins are unnecessary, and the 92-site test showed puppeteer-extra stealth at 72.3% pages served versus Fortress at 86.4% on the same harness. Removing the plugins also removes the JS-patch tells they can introduce.
What does Tilion Cloud cost?
Tilion Cloud is in beta with a waitlist, and rates are not published. Fortress v3 is free for developers with no caps; larger organizations need a production subscription, with terms at fortress.tilion.com/pricing. Do not compare the two products' pricing as if they were one.
What does Browserbase do better?
Stagehand and its SDK ecosystem, the Session Inspector debugging UI, four regions to Tilion's three, Selenium support, hosted Agents, the Model Gateway, x402 pay-per-use, and SOC 2 Type II with a HIPAA BAA on its top self-serve tier. Browserbase also completed 100 of 100 Browser Arena runs on 22 October 2026 where Tilion completed 99.
Can I try Fortress before committing?
Yes. Run docker run --rm -p 9222:9222 tilion/fortress:latest and connect Puppeteer to port 9222, or pip install tilion-fortress and launch locally. Without activation it runs on the public v1 engine; tilion activate unlocks v3 with an offline license check.
If protected sites are blocking your agents, join the Tilion Cloud waitlist at tilion.com and run Fortress locally today.