Product updates, engineering deep dives, and field notes from teams running real workloads on Kimchi.
Every vendor gets the security questionnaire, and most AI coding vendors answer it with a trust portal link. Here are the twelve questions security actually sends, answered in plain language you can paste into the response unedited.
Read article"Self-hosted" hides three different boundaries: execution, agent, and inference. At enterprise scale, the ability to move the inference boundary without rewriting the application becomes a first-class property of an AI development platform.
Platform teams rarely win the argument about which agent engineers use, and they don't need to. Cost and control don't live in the agent, they live in the API path underneath it. Two lines and one key, and every request gets per-user attribution and a budget that holds.
Same model, same benchmark, same conditions. Kimchi passed 69.4% of Terminal-Bench tasks. DeepSeek Harness's best preset passed 47.1%.
Safety systems never arrive before a technology - they arrive when it's already in use. Agentic AI is following the same pattern, and governance can't be bolted on once agents are in production. It has to be built in from day one.
A $500M Claude bill sounds like a freak accident. It wasn't. The controls to stop it existed the whole time - they just didn't enforce. Alerts notify. Caps enforce. Here's what governance that actually holds looks like.
We ran GLM-5.2 through four AI coding harnesses on Terminal-Bench 2.1. Same model, same 89 tasks, same benchmark. The harness made a 13-point difference. Here's what that means.
Real breakpoints, real stack traces, live variable inspection. No more print statements. Kimchi speaks the Debug Adapter Protocol natively, so the agent observes runtime state instead of guessing at it.
Seven countries have restricted the DeepSeek app on government devices. Zero have restricted the DeepSeek model weights. Here's why the distinction matters for your security review.
AI coding is becoming the software factory. Enterprises need to own the enforcement path, the economics, and the definition of done.
Open weights now sit within 1-3 points of the frontier on individual coding benchmarks at 6-30x lower cost. On composites, private benchmarks, and long-horizon agentic work, the frontier lead is intact.
~120,700 requests. 1,580 active sessions. ~98% success rate. ~$4.4k total spend - about 1/13th of the modeled frontier-model cost.
Why utility-priced AI coding tools blow budgets when deployed like SaaS, and how Kimchi's model orchestration and spend console fix it.
Today we take Kimchi Coding out of Early Access. It's live at kimchi.dev with full production support behind it — and this is the story of why we built it.
How Kimchi routes seven coding roles across Kimi K2.7, MiniMax M3, and Nemotron-3-Ultra - and why matching the model to the task cuts cost 40–60% without sacrificing quality.
The difference between autocomplete tools and a coding harness is not the model. It's the architecture.
We added MiniMax M3 to the recipe. Kimchi Coding just got spicier!
Why Generative AI is Breaking the SaaS Procurement Model
Kimchi is ready for you to try. Describe what you need, approve a plan, and the agent builds through every milestone on its own — until there's a PR waiting for you. Not a question. A PR.
Install Kimchi, connect your coding tools, and keep control from the first request.