Mozilla Announces 'TABS API' For Developers Building AI Agents

"Fresh from announcing it is building an AI browsing mode in Firefox and laying the groundwork for agentic interactions in the Firefox 145 release, the corp arm of Mozilla is now flexing its AI muscles in the direction of those more likely to care," writes the blog OMG Ubuntu: If you're a developer building AI agents, you can sign up to get early access to Mozilla's TABS API, a "powerful web content extraction and transformation toolkit designed specifically for AI agent builders"... The TABS API enables devs to create agents to automate web interactions, like clicking, scrolling, searching, and submitting forms "just like a human". Real-time feedback and adaptive behaviours will, Mozilla say, offer "full control of the web, without the complexity."

As TABS is not powered by a Mozilla-backed LLM you'll need to connect it to your choice of third-party LLM for any relevant processing... Developers get 1,000 requests monthly on the free tier, which seems reasonable for prototyping personal projects. Complex agentic workloads may require more. Though pricing is yet to be locked in, the TABS API website suggests it'll cost ~$5 per 1000 requests. Paid plans will offer additional features too, like lower latency and, somewhat ironically, CAPTCHA solving so AI can 'prove' it's not a robot on pages gated to prevent automated activities.

Google, OpenAI, and other major AI vendors offer their own agentic APIs. Mozilla is pitching up late, but it plans to play differently. It touts a "strong focus on data minimisation and security", with scraped data treated ephemerally — i.e., not kept. As a distinction, that matters. AI agents can be given complex online tasks that involve all sorts of personal or sensitive data being fetched and worked with.... If you're minded to make one, perhaps without a motivation to asset-strip the common good, Mozilla's TABS API look like a solid place to start.

Yet, no Scapbook

[BrendaEM]

Scrapbook was one of the best Internet research tools, which let the user save entire web pages, with depth, and highlight and add notes.. While there have been attempts to recreate it--a programmer is no longer allowed to make such a powerful extensions, such as XUL provided--and now, the AI proveyers have peddled their influence on Firefox, which will launch Firefox further into creepy spyware territory.
https://github.com/danny0838/firefox-scrapbook

Re:

[test321]

Scrapbook is the reason I keep a Palemoon install even though I use Firefox.

Re:Yet, no Sc[r]apbook

[shanen]

Hmm... Does that sound like a feature I would want to help pay for? The answer may surprise me. I think it sounds like a "Maybe" or even a "Yes" if the description was fleshed out a little bit. However I can also see where it belongs in an optional category for people who want it... Seems to me like the real cost would be quite large, but for an "ancient" and kind of fundamental reason: The HTTP links only go one way. That means there's no easy way for Scrapbook to know the target webpage has changed...

So a

Re:

[BrendaEM]

Scrapbook ran fine on older XUL Firefox, and continues to run under Basilisk. When I loaded my old data, which is some 4,000 web pages in several branches of science, I realized when I had lost when Firefox was XUL-nerfed. I tried to post something here on Slashdot about Scrapbook, but like most of my submissions, it was ignored. Scrapbook was so useful that Mozilla came out with its attempted equivalent competitor Pocket as a service, but the old Scrapbook users perhaps did not want the privilege of paying

"connect it to your choice"

[oldgraybeard]

or not!

Missing the point

[nategasser]

We don't need AI agents clicking and scrolling for us. That's like building a robot to drive a horse to take us to the store, when what we really want is to shop from home.

And we definitely don't need agents solving CAPTCHAs. That's just proof we have no idea what's going on.

We need websites that offer services (banks, travel booking, movie tickets, whatever) to adopt a communication standard (MCP seems to be the only game in town so far) so our agents can do these tasks for us in a way thatâ(TM)s reli