kimchi.dev (@getkimchi) header
kimchi.dev (@getkimchi) profile picture

kimchi.dev

@getkimchi

A CLI coding agent, inference layer, and spend console all in one. Powered by open-source models.

 United States
kimchi.dev
Joined April 2026
  • Tweets55
  • Following6
  • Followers245
14Media
Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)Photo by kimchi.dev (@getkimchi)
AI-generated code is only the beginning. At KubeAuto Day, Žilvinas Urbonas explores how Hermes + Kimchi fit into continuous, AI-native engineering workflows. Nov 9 · Salt Lake City · Free kubeauto.day/nal
1
250
Still looking for the perfect plan for the day before KubeCon? 🪐 KubeAuto Day is free, no KubeCon ticket needed. Kelsey Hightower opens the day, with Liz Rice, Bret Fisher and 12 more, in the IMAX and Dome theaters at Clark Planetarium, a short walk from KubeCon. Our VP of Engineering Žilvinas Urbonas shows why coding agents alone aren’t enough, and how Hermes + Kimchi make AI-native engineering the default. iPhone Duo and 15 GTA VI licenses up for grabs. Lunch, dinner and drinks on us. Nov 9, Salt Lake City. Register 👇
Speaker card for KubeAuto Day North America, November 9. “I will be speaking: Hermes + Kimchi = AI-Native Engineering by Default.” Žilvinas Urbonas, Vice President of Engineering, Kimchi, pictured in a black t-shirt on a blue background. kubeauto.day/north-america

ALT Speaker card for KubeAuto Day North America, November 9. “I will be speaking: Hermes + Kimchi = AI-Native Engineering by Default.” Žilvinas Urbonas, Vice President of Engineering, Kimchi, pictured in a black t-shirt on a blue background. kubeauto.day/north-america

2
3
179
No KubeCon ticket needed, the day before #KubeCon 👉 luma.com/ktykjzbj?utm_source…
Kubernetes Automation Day: Free, IMAX Dome, Liz Rice, Kelsey Hightower, Bret Fisher + 11 more...

Kubernetes Automation Day: Free, IMAX Dome, Liz Rice, Kelsey Hightower, Bret Fisher + 11 more...

Your KubeCon week starts here. For free. 15 leading cloud-native practitioners. Two theaters. Hands-on workshops. An iPhone Duo and 15 GTA VI licenses up for…

luma.com
1
31
Every request from an AI coding agent sends code and context somewhere. For most teams, that’s outside their network. What can go to a hosted model, and what should stay self-hosted? Lachezar Tsonov wrote up where to draw the line.
Diagram showing where each part of an AI coding agent runs: the tools, the agent loop and the model. Each can run on your own infrastructure or in a vendor’s cloud. When the model runs in the vendor’s cloud, code and context leave your network.

ALT Diagram showing where each part of an AI coding agent runs: the tools, the agent loop and the model. Each can run on your own infrastructure or in a vendor’s cloud. When the model runs in the vendor’s cloud, code and context leave your network.

2
1
74
Full piece here: kimchi.dev/blog/dynamic-infe…
36
Unplug the internet. Can your AI coding agent still do its job?“Self-hosted” usually means your tools run locally. The agent loop and the model often don’t. If the model runs in a vendor’s cloud, your code and context leave your network. kimchi.dev/blog/dynamic-infe…
Table titled “Where each part runs,” comparing where execution, the agent loop and inference run for four setups. Cursor self-hosted workers: execution on customer infrastructure, agent loop and inference in the vendor cloud. GitHub Copilot with data residency: inference with the vendor, in-region. Kimchi Serverless: local harness, local loop, inference on Kimchi GPUs with regional endpoints. Kimchi Self-hosted: local harness, local loop, inference in the customer’s Kubernetes. Below: “The disconnect-my-internet test: remove all access to outside networks. Can the application, agent and model still perform the required workflow?”

ALT Table titled “Where each part runs,” comparing where execution, the agent loop and inference run for four setups. Cursor self-hosted workers: execution on customer infrastructure, agent loop and inference in the vendor cloud. GitHub Copilot with data residency: inference with the vendor, in-region. Kimchi Serverless: local harness, local loop, inference on Kimchi GPUs with regional endpoints. Kimchi Self-hosted: local harness, local loop, inference in the customer’s Kubernetes. Below: “The disconnect-my-internet test: remove all access to outside networks. Can the application, agent and model still perform the required workflow?”

1
1
67
239k GitHub stars in seven weeks, and 22 points behind Kimchi on the same model. We ran DeepSeek Harness on Terminal-Bench with GLM 5.3 Flash: dsh's best preset passed 47.1% of tasks, Kimchi 69.4%. Results: kimchi.dev/blog/deepseek-har…
Benchmark results card comparing DeepSeek Harness and Kimchi on the same model and benchmark. DeepSeek Harness, best preset (creator): 47.1% task pass rate, total run cost $27.50. Kimchi: 69.4%, total run cost $36.03. Kimchi leads by 22.3 percentage points. Headline: the harness makes the difference. Terminal-Bench, GLM 5.3 Flash, maximum reasoning, October 2026.

ALT Benchmark results card comparing DeepSeek Harness and Kimchi on the same model and benchmark. DeepSeek Harness, best preset (creator): 47.1% task pass rate, total run cost $27.50. Kimchi: 69.4%, total run cost $36.03. Kimchi leads by 22.3 percentage points. Headline: the harness makes the difference. Terminal-Bench, GLM 5.3 Flash, maximum reasoning, October 2026.

1
1
139
more replies
We used GLM 5.3 Flash because the goal was to compare harnesses. If the model changes too, you can't tell which one made the difference.
1
36
Setup: official Terminal-Bench methodology, five trials per run, max reasoning, October 2026. Full writeup: kimchi.dev/blog/deepseek-har…
We ran DeepSeek Harness against Kimchi on the same model. It lost by 22 points | Kimchi Blog

We ran DeepSeek Harness against Kimchi on the same model. It lost by 22 points | Kimchi Blog

Same model, same benchmark, same conditions. Kimchi passed 69.4% of Terminal-Bench tasks. DeepSeek Harness's best preset passed 47.1%.

kimchi.dev
27
We used GLM 5.3 Flash rather than a DeepSeek model because the goal was to compare harnesses. If the model changes too, you can't tell which one made the difference.
1
19
Setup: official Terminal-Bench methodology, five trials per run, max reasoning, October 2026. Full writeup: kimchi.dev/blog/deepseek-har…
We ran DeepSeek Harness against Kimchi on the same model. It lost by 22 points | Kimchi Blog

We ran DeepSeek Harness against Kimchi on the same model. It lost by 22 points | Kimchi Blog

Same model, same benchmark, same conditions. Kimchi passed 69.4% of Terminal-Bench tasks. DeepSeek Harness's best preset passed 47.1%.

kimchi.dev
16
How do you govern a coding agent you can’t predict? @kunalD_official answered that at @Analyticsindiam Cypher, India’s top AI summit. Our Field CTO Kim-Norman asks the next question: what if every agent request sends your code to a US cloud? KubeAuto Day Berlin, Oct. 15 🌶️
kunal das (@kunalD_official) profile picturekunal das@kunalD_officialOct 13
Spoke at @Analyticsindiam Cypher 2026, India's biggest AI summit on how do you govern a coding agent you can't predict? You can't predict it. You can bound what it's allowed to do. That's the whole talk. @getkimchi is going places! #Cypher2026 #AIAgents
Image from kunal das (@kunalD_official)
Image from kunal das (@kunalD_official)
Image from kunal das (@kunalD_official)
Image from kunal das (@kunalD_official)
1
2
117
Free to attend, last spots here 👇 luma.com/kubeauto-berlin?utm…
KubeAuto Day Berlin · Luma

KubeAuto Day Berlin · Luma

KubeAuto Day Berlin is part of the global Kubernetes Automation Day series, bringing the cloud-native crowd in Germany together for an evening of Kubernetes…

luma.com
27
Same AI model, two coding agents. Kimchi solved 69.4% of Terminal-Bench tasks. DeepSeek Harness solved 47.1%. Surprise: DeepSeek's bare-bones setup (2 tools) scored higher than its full setup. Both ran on GLM 5.3 Flash. kimchi.dev/blog/deepseek-har…
Benchmark results card. Same model, same benchmark: the harness makes the difference. Kimchi: 69.4% task pass rate, total run cost $36.03. DeepSeek Harness, best preset (creator): 47.1%, total run cost $27.50. Kimchi leads by 22.3 percentage points. Terminal-Bench, GLM 5.3 Flash, maximum reasoning, October 2026.

ALT Benchmark results card. Same model, same benchmark: the harness makes the difference. Kimchi: 69.4% task pass rate, total run cost $36.03. DeepSeek Harness, best preset (creator): 47.1%, total run cost $27.50. Kimchi leads by 22.3 percentage points. Terminal-Bench, GLM 5.3 Flash, maximum reasoning, October 2026.

2
6
297
Our table is set and the agents (and speakers) are cooking 🌶️ Come say hi at KubeAuto Day Paris and grab some spicy stickers.
Cast AI (@cast_ai) profile pictureCast AI
@cast_aiOct 8
KubeAuto Day Paris is live 🇫🇷 From an empty room this morning to a full house at La Bellevilloise. Talks on Kubernetes and AI-driven automation are underway. Live updates from every talk in the thread 🧵👇
1
39
Your LLM bill doubled and nobody knows which team caused it? 🗽 The Kimchi team is coming to @aiDotEngineerNYC to help. Find us at booth 17, Oct 12 to 14, Sheraton Times Square. Model routing, AI costs and governance: come talk to us 👋
Kimchi promotional card on a dark New York skyline background with the Brooklyn Bridge. Text reads: "Meet Kimchi team at AI Engineer NYC. October 12 to 14, 2026. Booth 17. Sheraton New York Times Square Hotel." Kimchi and AI Engineer New York logos.

ALT Kimchi promotional card on a dark New York skyline background with the Brooklyn Bridge. Text reads: "Meet Kimchi team at AI Engineer NYC. October 12 to 14, 2026. Booth 17. Sheraton New York Times Square Hotel." Kimchi and AI Engineer New York logos.

4
79
Cars came before seatbelts. Technology spreads first, safety catches up later. Today it's AI agents with your repo, your credentials, your terminal. Nobody hands a new hire a master key and a polite note. Governance can't be bolted on: kimchi.dev/blog/why-agentic-…
1
39
96
$500M on Claude, and the controls to stop it existed the whole time. Alerts notify. Caps enforce. Most AI coding tools ship the first and call it governance. Ask your vendor: when a team hits budget, does spend stop, or do you just get a ping? kimchi.dev/blog/the-500m-bil…
1
39
Load more