Learning Library

← Back to Library

Manis AI: Claude Sonnet with 30 Tools

Key Points

  • Manis AI, presented by a Chinese startup, debuted with a demo managing dozens of social media accounts, but was later revealed to be Claude Sonnet augmented with about 30 integrated tools rather than a brand‑new model.
  • The system can generate highly detailed outputs—comparable to GPT‑4 and Deep Research—but suffers from scaling problems such as slow response times and occasional errors as the team works to secure enough hardware.
  • Legal questions arise regarding whether repackaging Claude under a new name for commercial use violates its terms of service and how enforceable those restrictions are under Chinese jurisdiction.
  • The episode highlights a broader trend: existing AI models can achieve remarkable new capabilities simply by adding better tooling, suggesting that the upcoming “agent age” may focus more on tool integration than on developing entirely new intelligences.

Full Transcript

# Manis AI: Claude Sonnet with 30 Tools **Source:** [https://www.youtube.com/watch?v=82ifRQrwEfM](https://www.youtube.com/watch?v=82ifRQrwEfM) **Duration:** 00:03:54 ## Summary - Manis AI, presented by a Chinese startup, debuted with a demo managing dozens of social media accounts, but was later revealed to be Claude Sonnet augmented with about 30 integrated tools rather than a brand‑new model. - The system can generate highly detailed outputs—comparable to GPT‑4 and Deep Research—but suffers from scaling problems such as slow response times and occasional errors as the team works to secure enough hardware. - Legal questions arise regarding whether repackaging Claude under a new name for commercial use violates its terms of service and how enforceable those restrictions are under Chinese jurisdiction. - The episode highlights a broader trend: existing AI models can achieve remarkable new capabilities simply by adding better tooling, suggesting that the upcoming “agent age” may focus more on tool integration than on developing entirely new intelligences. ## Sections - [00:00:00](https://www.youtube.com/watch?v=82ifRQrwEfM&t=0s) **Manis AI: Claude Wrapper Unveiled** - Manis AI, a Chinese startup’s tool launched over the weekend, was revealed to be a sophisticated wrapper around Anthropic’s Claude Sonnet equipped with dozens of integrated tools rather than a brand‑new model as initially rumored. ## Full Transcript
0:00news moves fast over the weekend Manis 0:04AI became a huge thing across the AI 0:07world and I'm going to explain you what 0:08it is basically Manis is like if deep 0:11research was crossed with claw and 0:15that's not an accident because later in 0:17the weekend so this all came out Friday 0:19right men has dropped their release 0:21video they're a Chinese startup they 0:24came out and shared a lot about sort of 0:25what they're accomplishing and why 0:27they're doing it and their p passion for 0:30sort of making this more available to 0:32people they had a killer demo video 0:34where I think they were managing like 26 0:35social Accounts at once using 0:38Manis and I thought well we'll have to 0:41see how this goes right um well it turns 0:46out that Manis is a really fancy rapper 0:50for Claude and that came out by Sunday 0:54and so someone figured out how to drop 0:55out sort of the tool set for Manis and 0:58what it looks like is that it's Claude 0:59Sonic 1:00with I think 30 different tools that it 1:03can 1:03call and the results are real like the 1:07the system is slow it takes a really 1:09long time sometimes it errors out which 1:12I think is just scaling issues for the 1:13team they're trying to get those chips 1:14in as quick as they 1:16can but it does work uh it produces uh 1:20highly detailed results uh stuff that's 1:23on par with 03 uh and deep 1:27research and everyone is super assumed 1:30at the beginning of the weekend that 1:31that was 1:32because it was a new model that the 1:36Chinese lab had built but it turns out 1:38it was Claude Sonet and what the team 1:40had done is actually agen toy Claude 1:42Sonet they'd given it this wide range of 1:44uh tool sets that that Claude could use 1:48they'd given it special instructions and 1:50they'd wrapped it up and that's 1:54Manis so we'll see how long that lasts 1:57um I'm not sure if the terms of service 1:59for Claude include commercializing 2:01Claude and calling it something else um 2:03and I'm not sure how enforceable that 2:05would be uh under Chinese jurisdiction 2:07for Claud anyway that's a question for 2:09the lawyers um and it is possible that 2:12more of the story will come out and 2:13people will say oh no no it wasn't you 2:15know it wasn't really claw they have 2:18something different um so we'll see it's 2:21it's an emerging conversation I think 2:24the think I I think the thing that is 2:26most interesting to me is that I was not 2:30surprised when I heard that it was 2:32clawed in 30 Tools in other 2:35words we have reached the point in the 2:38AI Revolution where if someone tells you 2:41I'm using an existing an existing 2:44intelligence but I gave it new tools and 2:46it can do something absolutely 2:47incredible it can create a 75 page 2:49document on a new medical treatment for 2:52multiple sclerosis which was one of the 2:54things that it actually did over the 2:56weekend um I'm not surprised 3:00I'm no surprise and that's where we're 3:03at and it underscores to me that we talk 3:05a lot here about how quickly 3:08intelligence is scaling and rightly 3:11so but what we don't talk about as much 3:13is the intelligence we already have can 3:15go a long way with a little bit better 3:18tooling and I wonder if part of the 3:20agent age in 2025 is going to be 3:24essentially taking the intelligence that 3:25we've got which is already very good and 3:27putting wheels on it right giving it 3:29tooling giving it the chance to actually 3:31go out and get something done so we will 3:34see how the story of Madness unfolds a 3:36bit more uh I'll probably be writing 3:38more on it talking more about it it's a 3:40big story uh it's very very difficult to 3:42get performance like deep research 3:44congrats to the team however they did it 3:46um it's an achievement regardless and 3:49sort of we'll see where we go from here 3:50good luck with the scaling guys