85,000 GitHub stars on RTK.
In January: one Rust binary that cleaned up terminal output for AI agents. Since August: RTK Pro, governance and security for teams, now used by our first customers.
Thank you to every contributor and every early customer.
"Discord Drops Electron for Rust Rebuild, Early Windows Build Sips 290MB of RAM"
it seems the big exodus has started . nice.
On rtk oss and prod we already full rust.
@BoboLinNotes Agreed. Even self-hosted, Revolut capped context at 256k. The average Claude Code request is already ~84k tokens (Requesty data) and Anthropic says context per request grew 2.6x in six months. The goal is agents that work well at 256k, with less noise in the window.
Revolut moved ~2,000 engineers off Claude Code and Codex to self-hosted GLM on 56 B200s. Bill ~10x lower.
The real lesson: their gains came once token limits were lifted. Agents work better when you stop rationing token
That’s what RTK cuts, open source, before the agent reads it. RTK Pro shows where the tokens go, per team and project.
Revolut’s talk: https://t.co/HeIbvxoj9R
Most teams can’t self-host. They’ll pay per token, and plans will get pricier or tighter as agents burn more.
So the lever left is tokens per task. A lot of what agents read is noise: progress bars, passing tests, full logs, resent every turn.
85,000 GitHub stars on RTK.
In January: one Rust binary that cleaned up terminal output for AI agents. Since August: RTK Pro, governance and security for teams, now used by our first customers.
Thank you to every contributor and every early customer.
- Share a pipeline as one file. - Plug in new methods that every pipeline can use. - Agent agnostic pipelines: agent agnostic TOML pipeline definition.
Open source, written in Rust
github rtk-ai/rtk
-> RTK engine, beyond bash compression..
- Compress more than bash: lockfiles, search results, web pages, and the API traffic itself.
- Build your own pipeline in one file.
- Delegate sub tasks and actions by complexity level to cheaper models.
- Delegate sub tasks and actions by complexity level to cheaper models. - Share a pipeline as one file. - Plug in new methods that every pipeline can use. - Agent agnostic pipelines: agent agnostic TOML pipeline definition.
-> RTK engine, beyond bash compression..
- Compress more than bash: lockfiles, search results, web pages, and the API traffic itself.
- Build your own pipeline in one file.