kozh ./

Free AI tools + posts about AI & crypto

sharing useful finds, guides and opportunities

dms open

[kozh ./ @mr_kozh]

https://t.co/5FCWpy27cQ

[kozh ./ @mr_kozh]

@qwinsi0x The wallet that turned out well

[kozh ./ @mr_kozh]

@0xDepressionn

[kozh ./ @mr_kozh]

why pay $300/mo for Grok Bot
when the open-source version is already on GitHub?

Rakazo just crossed 1k stars in a week⭐️

Same idea:
• always-on AI teammates
• with their own persistent computer.

Only this one is free and you bring your own model.

[kozh ./ @mr_kozh]

https://t.co/5FCWpy27cQ

[kozh ./ @mr_kozh]

@0x_fokki Each one is useful

[kozh ./ @mr_kozh]

@Kontentsukpi I’ll go and have a look too

[kozh ./ @mr_kozh]

repo: https://t.co/Hdfqv7dbCg

[kozh ./ @mr_kozh]

@pazhik_ido @Collector_Crypt @Courtyard_io @phygitals @Beezie @mnstr @GEMINT @Grailed_GG @JupiterExchange @renaissxyz @gacha_game_ I haven't heard of this trend

[kozh ./ @mr_kozh]

@beamnxw Useful and valuable – I use it

[kozh ./ @mr_kozh]

@_tifa_67

[kozh ./ @mr_kozh]

@commonsmade hey @commonsmade i vouch for @Nikitont

[kozh ./ @mr_kozh]

Opus solved 74 tasks for $54.73.
This GitHub router solved 75 for $25.66.

UncommonRoute is a local model router for Claude Code, Cursor and Codex.

instead of sending every request to the strongest model, it decides how difficult the task actually is.

simple stuff goes to cheaper models.

medium tasks get a stronger one.

complex requests can still escalate to your best model.

in their SWE bench test:

• Opus only
74 / 100 solved
$54.73

• UncommonRoute
75 / 100 solved
$25.66

53% less API spend.

the useful part is that routing happens for every request individually.

so fixing a typo and designing a distributed system don't have to cost the same.

https://t.co/UiZtqAtD2e

[kozh ./ @mr_kozh]

@me_cool_off Do you use something like that?

[kozh ./ @mr_kozh]

Claude Code can waste thousands of tokens on logs and JSON. This GitHub repo cuts that before the model even sees it.

I found a GitHub project called Headroom.

the problem is pretty simple:

coding agents constantly read logs, JSON, tool outputs, files and RAG results

and a lot of that context is just noise.

Headroom sits between the tool output and the model, compresses the unnecessary parts, then sends the smaller version to Claude, Codex, Cursor or whatever agent you're using.

their benchmarks show:

- 15 to 20% fewer tokens for coding agents
- 60 to 95% fewer tokens for JSON
- 17,765 tokens to 1,408 on code search
- and it can run as a proxy, library or MCP server.

the part I like is that you don't need to change how the agent works.

you just reduce how much garbage it has to read.

https://t.co/B6hW8Y04Fv

[kozh ./ @mr_kozh]

https://t.co/5FCWpy27cQ

[kozh ./ @mr_kozh]

@0x_fokki

[kozh ./ @mr_kozh]

@antpalkin Have a go – I hope it helps

[kozh ./ @mr_kozh]

@beamnxw thanks

[kozh ./ @mr_kozh]

@0xMfox I don’t know why this isn’t included in the standard models yet

[kozh ./ @mr_kozh]

@0x_fokki There are emotions, but you can still tell it’s AI

[kozh ./ @mr_kozh]

@antpalkin I’ll save it so I can have a look at it when I’ve got a spare moment

[kozh ./ @mr_kozh]

@0xwhrrari the sequence of actions is correct as he interprets it

[kozh ./ @mr_kozh]

@me_cool_off @variational_io Congratulations on this achievement