>>109481014
>The data sent back and forth through the provider's API has to have a structured output.
That’s not MCP-specific, that’s literally how every API ever made works, what the fuck are you smoking?

guy who doesn't understand how computers work continues to insist he's right

>oh no, it went up a couple of pennies. it's dead
china's shills really are a bunch of third worlders who live in villages with a median household income of $500 lmao

File: file.png (1.99 MB, 1448x1086)

1.99 MB PNG

>August 3rd - Alibaba releases Qwen3.8-Max
2.4T total / ~95B active MoE, 1M context, multimodal. Hosted live ($2/$6). Independent scores in (Artificial Analysis Intelligence Index 56, strong agentic/GDPval numbers). Open weights (plus a 27B companion) still listed as coming the week after launch — not confirmed out as of midday Aug 6.
>July 31st - DeepSeek promotes V4-Flash-0731 to production
Same 284B/13B active MoE + 1M context as the preview, major post-training gains on agent/coding benches. Beats their own larger V4-Pro-Preview on published numbers. Silent API upgrade, same low pricing, MIT weights.
>July 30th - OpenAI cuts GPT-5.6 Luna 80% and Terra 20%
Luna to $0.20/$1.20, Terra to $2/$12. Sol unchanged but gets Fast mode. Efficiency gains + competitive pressure from cheaper options.
>July 27th - Alibaba quietly launches Qwen3.7-Flash
Cheap 1M-context multimodal aimed at high-volume agent/vision workloads ($0.03/$0.13).
>July 26/27th - Moonshot drops full open weights for Kimi K3
2.8T-param MoE, largest open-weight model at the time, 1M context, multimodal.
>July 24th - Anthropic releases Claude Opus 5
Near-Fable 5 performance across many categories at half the price ($5/$25, Fast mode available). New default on Claude Max, effort dial, 1M context.
No major new frontier model releases in the last day-plus. Side developments include Google DeepMind leadership changes (Hassabis moving to chair/Alphabet Chief Scientist role), Meta shipping Muse Code (terminal coding agent), and continued reports of containment/eval incidents across multiple labs.

AI made hardware unaffordable.
Fuck AI. Stop giving money to AI companies.

>>109481058
> filthy human thinks he can make the most out of shinny new hardware
do you have any idea on how many luddities have an RTX 5090 and only play Runescape all day long?

>>109481041
It's not MCP specific but if what you mean to convey is that MCPs are eating disproportionate usage for tool calls, then you should also consider that tools have to be called during normal usage even without a "MCP" connected. The model itself has to interact with an API. Through JSON payloads, for example. Like a MCP.

>>109481089
the guy you're talking to thinks mcp's should be magical bridges that enable communications between computer programs that cannot normally have programmatic communication
lol

>>109481051
What's shameful about being born in a certain part of the world?

>>109481051
>with a significant increase expected
but of course ai niggers cant read

>>109481094
lmao

SERIOUS QUESTION: which AI will model out a fully functioning woman's body with joints, flopping tits and pussy
I am willing to cash out 100 dollars for any AI that can do this

>>109481127
Stop shitting up the thread with your stupid requests.

>>109481089
>Like a MCP
This is the dumbest meme. You are describing the definition of API, an ancient definition, then trying to say APIs are “like MCP”, something brand new. What is MCP adding?
Name one actual thing you are using it for. There are use cases, “blender” (if not external and shell-inaccessible) and “documentation” are not those cases. Those are wasteful uses of MCP. This definition rewriting shit you’re trying to pull is beyond retarded.
You keep bringing up APIs and JSON which have existed forever and agents use these all the fucking time. What does MCP add besides the schema, something that can be instantly stripped out and used on its own without needing to change anything else?

>>109481127
How many times are you going to ask. Your options are
>just try it
>download one of the many existing assets
>kill yourself

>>109481127
Keep asking about your genius requests.

>Sol XHigh in charge of a bootstrapped self mutating thinking tool
hold me bros I’m scared

>>109481143
Thanks for the positive feedback, anon — I will ask again next thread.

>>109481212
No problem, chad.

>>109481141
i'll give you the benefit of the doubt one time and give you a concrete example of where an mcp would be useful and save you tokens
i had a local program with a very robust api and very easy to just send code to
the api is well documented but training data is sparse
some parts of the api only came into existence after the models had been trained because the program is constantly updated

there's a certain thing i ask the clanker to do in this program frequently - it fumbles over this despite the good docs and well documented api
this thing is not a single api call, but actually a sequence of calls that need to be run in a particular order for the intended effect to be achieved
this fumble happens every time wasting literally tens of thousands (close to 100k) each time

because i do this so often it's nice to just have a single tool to call instead
i don't use an mcp for this, but it's a case where you could use one and save a mountain of tokens

i'd explain how a skill can also help in such cases, but you only seem to use software that's covered by training data so you probably can't appreciate the idea of having to teach a model entirely new things inside the context window

Created a master list of gemm shapes from all the models in diffusers and transformers, I downloaded all the configs to do it. I’ll use the list for the profiler in my inference engine, I need to narrow down the gemm candidate list, im thinking some of them probably never win at any shape

File: i.png (76 KB, 631x1702)

76 KB PNG

>>109481096
>>109481141
There are some nuances obviously. But to me it seems like run-of-the-mill calls still consume tokens for both input and output, irrespective of the fact that there's an intermediary normalizing data. Pic related.

MCPs can bloat your context wildly because when you "activate" them in the session, the whole shebang gets dumped into context: actions, their descriptions, parameters, potentially thousands of them. It doesn't have to be that way, a plugin can have a narrow surface.

You don't need a plugin for local shell usage and it would be retarded to use one in that case. A plugin can be beneficial if instead of the model having to write its own scripts or huge bash one liners reliant on exit codes each time, it gets a simple interface with a few known parameters.

>>109481238
>this thing is not a single api call, but actually a sequence of calls that need to be run in a particular order for the intended effect to be achieved. this fumble happens every time wasting literally tens of thousands (close to 100k) each time
You are describing what a script is used for. I’ve watched Sol XHigh create python and typescript scripts to solve exactly this problem in my own tools. I used the session logs to make my tools better, but if you don’t have control over the tool, you just make a script.
You described a deterministic, repeated action. AI has absolutely nothing to do with that.
Your AI was fumbling because it’s poor quality, low intelligence, just like you, and both of you failed to realize that wrapping the tool in a script or fixing the tool would’ve solved your problem forever and would’ve even reduced token cost.
Instead, you used a bloated specification to verbally explain the idea of a script, and your LLM now wastes huge amounts of tokens because you STILL haven’t fixed the issue and you’re also using a bloated protocol.
You are RETARDED.

>>109481051
good morning saaaar!

>109481286
you are eastern

>>109481127
pussies dont flop retard

>>109481051
I have an annual income of 300$.
Just steal the tokens lmao.

>109481286
marge the jeets are in /vcg/ again

got memed like a good goy into doing 5x on codex instead of claude. sol is trash and anything below it is unusable

>>109480547
anons.... now that OpenAI has no incentive for free resets, are we fucked?

>>109481280
anon, i use a cli command for this
and yes it's running a script
it's a script it's run literally thousands of times
it's nice to have it packaged in a cli
i'm not going to have the clanker rediscover how to write the script, waste hundreds of thousands of tokens every single time am i? nor am i going to put it in a folder that i always run it from and then point to script like a mongoloid and be like 'use this please'
it's a deterministic command to work around a known model failure
it's nice to package it up into a neat reusable package
or are you an ascended master who literally writes every piece of software from scratch whenever he needs it?
did you write the browser you're using right now today?
because it sounds like you should be able to with your genius iq

>>109481298
This. It feels like a bag of sand when you touch it.

>>109481308
skill issue

>>109481245
Your pic rel is the closest, maybe defines a smaller part of MCP, but the protocol is still bloated.
>clanker has a bearer token or other auth token
>clanker has a search engine endpoint
>clanker can use curl and can do so very efficiently
>give your clanker some extra instructions if needed, but not absolutely necessary
>ask the clanker exactly what you would’ve asked through MCP
and at least half of your image is rendered waste. For something like blender, 80% would be waste, that’s where the 5x figure comes from. You still are wasting HALF your tokens at least

>>109481320
Save the script? Make your cli do the sequence from one command?

>>109481341
the cli command is the script you fucking imbecile
that's the point of the cli - it's a collection of scripts put together to form a program...a cli
you know what else can do that? an mcp tool.

File: file.png (50 KB, 699x256)

50 KB PNG

What's your least favourite syscall?

>>109481320
>rediscover how to write the script
Of course not, you write the script once and just give him the file name. Have a comment in the script that says “this is what this script is for” and that’s it, everything else is waste. If you don’t want to type that out each time, drop the file path in an AGENTS.md, it’s not that hard, that’s extremely old tech now
None of what you said indicated any reason MCP needed to be involved

>>109481308
we told you anon... Sol tends to bloat shit up. It's only decent when following a plan made by Fable.

Also, Opus 5 is basically unlimited on the 5X plan, so you fucked up bad

>>109481357
are you trolling or are you actually retarded?

>>109481366
>Sol tends to bloat shit up
I laughed at this

>>109481373
Nobody cares, fuck off.

>>109481362
or i can just have the cli on path and a skill that says 'use cli for this'

>>109481372
show me your scripts folder, bro

>>109481320
>nor am i going to put it in a folder that i always run it from and then point to script like a mongoloid and be like 'use this please'
So you’re wasting thousands of tokens by refusing to use directories/folders properly and de facto making your clanker know every single tool on your machine every time?
What the fuck?
What did you even mean by this?
>nor am i going to put it in a folder that i always run it from
If it’s not in a folder already, where the fuck is it? Lmfao, you don’t even know how folders work?
???

>>109481016
>didn't update the last thread link
>>109476795

Stop linking agdg in the OP. The AI spammer already killed that general for months and it's only finally recovering. Almost a year of daily non-stop "snailcat" and related spam and zero vibecoded games were posted. It is not a related general. What you're doing is called raiding and it's against the rules of this website.

>start here if you have $20 or so https://claude.com/product/claude-code

Advertising is against the rules of this website.

>>109481238
Agreed that the vast majority of MCPs are wasteful.
I had a use case for Kubernetes, instead of the clanker having to discover across sessions how everything is set up via SSH, it gets a plugin with 5 actions that allows it to spawn pods, connect to ephemeral environments, send and monitor long jobs and so on. Saves time and tokens (partly because there are no failures anymore, the calls succeed 100% of the time).

>>109481094
kek

>>109481398
>So you’re wasting thousands of tokens by refusing to use directories/folders properly and de facto making your clanker know every single tool on your machine every time?
no?
do you understand how computers work?

>>109481051
>it went up a couple of pennies
The announcement literally said the price is being doubled.

How much is the price going up by? Its moving in the wrong direction. Google has a free trial, I might use it to try out my RSVP apps. Otherwise I need to save up another $100 to test more of this shit. Going to take me to next year (I have other stuff I need to buy first). Can't use anything but Deepseek. If it goes up by more than 50%, I'm not going to use it.

>>109481432
... wouldn't that be a couple of pennies?

>>109481388
>a skill that says 'use cli for this'
I think that really does narrow it down. Instead of just being in the proper folder, you’re just blindly instructing an agent “wherever” on the machine to “just get it done” and handing him every single binary in your PATH that you’ve described via skills, which it then has to use through bloated MCP, and it has to guess what tool to use because you refuse to say “use `tool`”
That’s fucking incredible

>>109481434
It's still cheap even a 50% increase.
Also you can just use openrouter for cheaper prices.

How does everyone plan to afford this shit when bagholders aren't subsidising it? Most the AI companies need around a 4x price increase (from what I've heard) to break even. You're approaching its cheaper just to do it yourself price at that point.

>>109481430
>nor am i going to put it in a folder that i always run it from
No really, do elaborate. Where are your programs located?

>>109481440
No. The total expenditure is now 2x. Sloppers are finally beginning to be priced out because providers are tired of carrying the burden. Paypiggies continue to gaslight and pretend otherwise.

>>109481443
there's no mcp and the skill literally just says

>when doing this sort of work run `examplecli --help`

it's <200 toks for the full command discovery process and another couple of hundred to run the command read back the validation

this is not hard or complicated

>>109481457
you actually have compiled programs on your computer now?

>>109481357
And why can’t you just make the script do what you want

>>109481462
You seem extremely emotional about this subject.

>>109481016
The deepseek API was costing me more than opencode go. If it goes up its basically not usable. I would be able to use opencode like an hour a day, or pay $100/month through the API to use it for like 4-5 hours a day. I'm just making very basic crud apps. Nothing even remotely interesting or particularly difficult.

>>109481413
The ban dodging spammer is the one who defaced the OP and anons are too lazy to fix it. This general has routinely complained about him making bad OPs early so he can post his gay forced memecat.

>there's no way to do it this way
>are you sure?
>that's a fair push. i took the lowest-effort route. let me investigate how to do it properly
Why has Claude become so lazy? Is Fable like that or just Opus?

>>109481470
>there's no mcp
Then what the fuck are you arguing with me about?
I’ve been saying MCP is almost always token waste and why that is, over and over again. Fuck off you illiterate retard.

File: ds4.png (193 KB, 949x990)

193 KB PNG

>>109481456
Hopefully by that time we'll have at least Deepseek Flash level intelligence at home. In the meantime I'm trying to do my part.

>>109481505
because you can do literally the same fucking thing with an MCP you dipshit

>>109481498
It's AGI.
Why should it do what you tell it to?

>>109481495
what about qwen flash?

>>109481489
Dude is two seconds from a mental breakdown

>>109481514
>i don’t use MCP personally
>and you don’t use it either
>you even say it’s wasteful
>but because it could be used
>even though neither of us use it
>probably because it’s wasteful
>I’m going to attack you
lmfao, it’s a joy to see your train of thought. Gemini could out-think you

>>109481016
>try deepseek api for the first time yesterday
>this happens
are they trying to stop me?

>>109481498
AI will always have some lazyness trained into it, otherwise it would just think for 5 hours to solve 1+1

File: 1757118618413571.png (1.17 MB, 864x1184)

1.17 MB PNG

I need a reset man.

File: file.png (58 KB, 1455x502)

58 KB PNG

>>109481549
If you don't use MCP you are a dumbass
https://github.com/runit-xze/unix-docs-mcp/

File: opencode-go.png (26 KB, 854x589)

26 KB PNG

>>109481495
Really? I'm finding the limits very hard to hit.
Post this chart, I wonder what yours looks like.
Are you breaking the cache somehow maybe?

>>109481567
it should be possible to sell my limits to other anons, I still have 60% and it resets in 2 days, and I don't know what to do with it.

I would sell you 40% of my 20X plan for $30

>tripfag retard uses MCP
no surprises there

>>109481549
i accept your concession
i went with cli over mcp at the time because context bloat via mcp was a problem then
there are ways to work around that now
hell, i wouldn't be surprised if even the most anti-MCP harness dev around (the guy who makes pi), ends up adding MCP support in the next few months
https://blog.cloudflare.com/mcp-v2/
https://blog.modelcontextprotocol.io/posts/2026-07-28/

>>109481366
I'm actually so upset at myself for not doing that. $100 down the fakkin drain

>>109481591
>there are ways to work around that now
Okay? And?
>does MCP magically save tokens? no? so I shouldn’t use it
>at best, it doesn’t cost that many more tokens! okay, so why should I refactor and wrap everything to use it?
You only add it if it’s exact super niche combo of requirements all apply to whatever you’re working on, you pay for the convenience of that through the bloat

>>109481628
What makes your Cloudflare product better than Lightpanda, PinchTab, agent-browser, Playwright Agent, Browserless, Kuri, ABP, chrome-cdp, OpenCLI, ego-lite, Browser4, and the dozens of others that come out daily?

>>109481628
give me agentic Safari usage and I will hand over so much money so that I never have to manually test one of these cursed fucking iPhones by hand ever again

>>109481628
Fucking pieces of shit.

>>109481671
Bro, come on.
YOUR browser wont be able to pass their "turnstile". Theirs will because 99% of the Internet is "protected" by them.

>>109481715
Cloudflare strong-arming people? No, couldn't be, no way. That wouldn't happen. Nope.

>>109481643
NTA but the best thing to if you see your frontier model struggling over time with something, then figuring it out, then struggling again in a separate session, or after repeated compaction, is literally just asking
>would a plugin help you waste less tokens on trial and error? consider schema bloat and read through https://developers.openai.com/api/docs/guides/tools-connectors-mcp#defer-loading-tools-in-an-mcp-server first

File: file.png (165 KB, 1502x715)

165 KB PNG

You know what else has no cow? Oatmilk. You should try drinking oatmilk!

File: kekdisto.png (117 KB, 280x512)

117 KB PNG

>>109481628
>KikeSurf :tm:
>Target audience: Jewish CEO numba 1 - sistersmasher
>Target audience: Jewish CEO numba 2 - wallstreet rat
>Target audience: Jewish CEO numba 3 - rocket jew
lol. lmao even

File: file.png (27 KB, 1486x123)

27 KB PNG

>>109481744
whoops, the story of the diverge of 4.2BSD vs Linux re: process forking got cut off.

File: media_HLFNyhsagAAVl9p.jpg (82 KB, 1280x720)

82 KB JPG

There's like 30 v4flash 0731 providers on openrouter. Can you faggots just chill out and choose one of the other providers so the whale can finally launch nu-v4pro?

Its kind of retarded we can't even see the flagship model because demand for the small-fast-cheap model is so high. Huawei needs to start pumping out superpods already

>>109481628
> kikesurf

>>109481759
DSv4flash is basically Opus4.6 tier and I was happily using that as SOTA half a year ago
Progress has genuinely been insane

File: lolplanning.png (14 KB, 687x597)

14 KB PNG

>>109481582
Most of it was Kimi. If I only use deepseek, I would be good for 4-5 days before the price increase. I have to cut it off after the reset in a few days. I only use Kimi for the planning and advising and deepseek for everything else. Its 97% deepseek after the initial kimi calls, except for occasional screenshots to verify the UI. The strange part is it did this yesterday too with the usage being really high in the morning. I'm only going to use deepseek from now on and then Luna for vision. The two different model strat is still too expensive.

>>109481737
>if you see your frontier model struggling over time with something, then figuring it out, then struggling again in a separate session, or after repeated compaction
…and we’re back to scripts and an AGENTS.md or really any written reminder.
If you have an agent struggling with a CLI, you can just make a new CLI that wraps the old one and has a better —help interface and a small bundle of docs for common issues, you can fork the existing CLI if it’s open source and just have your agent fix it, and again MCP is adding nothing, it doesn’t make your agents “smarter” in any way that’s not easily done without MCP
if your agent is struggling with a tool, your agent is too dumb and/or your tool sucks and not even MCP claims to be some magic fucking bandaid

File: 1762842962664672.png (2.07 MB, 1402x1122)

2.07 MB PNG

Does anyone use OpenCode go? How is it? Seems pretty nice value wise, but are their providers shit?

>inb4 I should OpenCode go fuck myself

>>109481801
The point is that it can evaluate its own performance and reason about tool usage better than the user. Maybe struggle isn't the term, it's all kinda relative now. A frontier model will reason through most things you can throw at it. But it can also write abstractions to help itself in the future and become more efficient. Whether they're scripts, instructions, automations, MCP, it doesn't matter that much.