What does the app do?
* Monitors/Records/Replays shell accesses, be it SSH shell, or k8s pod debug.
For video demo, https://vimeo.com/1151047130
What is the role of AI?
* This app is 100% developed by AI. Anthropic Opus 4.5 with Claude code cli.
* Spent roughly $50 for claude AI.
* Developed it in < 3weeks
* Frontend + Backend + Storage layer + Ephemeral Container image + SSH Shell recorder is all developed using prompts.
* The recording happens
Some highlights of AI:
* Claude reverse engineered the scriptreplay code to create a console player in react.
* Claude handle SSO integration and decided most of the
How does it work internally?
* For the SSH shell, it updates the SSH config such that on every login, a banner is displayed and a terminal recorder is started. On exit from the shell, the recorded file is sent to an S3 bucket from where the console replays it.
* For K8s Pods access, it uses ephemeral containers (available since k8s v1.25). The ephemeral container starts the terminal recorder and on exit, sends the recorded file to the s3 bucket. Admission controller is used to deny access to k8s pods using kubectl exec. Also Admission controller is used to deny use of any other ephemeral containers apart from the ones provided.
Can it be deployed onprem?
* Yes, all the components are chosen such that it can be deployed onprem in air-gapped environments too.
Why did I work on it?
* My aim with this exercise was to actually build a realistic tool that can help real users... There are a million other case studies where you could find AI helping users with one-off tasks. One never gets things right in the first iteration when building anything at scale. I assumed there would be rework, debugging, deprecating existing modules, and designing different modules. For scale and multi-tenancy, I changed the storage model multiple times. I wanted to check if AI would actually be effective in all these phases and what kind of issues I would face... working on an article explaining the challenges, cost, best practices for developing an AI-assisted app, and more importantly, what security practices could be followed while doing so. Note: I ran security scans on the auto-generated code and manually verified the findings. Then I used Claude itself to fix those security findings. To summarize, I believe AI could independently orchestrate full-blown enterprise apps.
What does the app do? * Monitors/Records/Replays shell accesses, be it SSH shell, or k8s pod debug. For video demo, https://vimeo.com/1151047130
What is the role of AI? * This app is 100% developed by AI. Anthropic Opus 4.5 with Claude code cli. * Spent roughly $50 for claude AI. * Developed it in < 3weeks * Frontend + Backend + Storage layer + Ephemeral Container image + SSH Shell recorder is all developed using prompts. * The recording happens
Some highlights of AI: * Claude reverse engineered the scriptreplay code to create a console player in react. * Claude handle SSO integration and decided most of the
How does it work internally? * For the SSH shell, it updates the SSH config such that on every login, a banner is displayed and a terminal recorder is started. On exit from the shell, the recorded file is sent to an S3 bucket from where the console replays it. * For K8s Pods access, it uses ephemeral containers (available since k8s v1.25). The ephemeral container starts the terminal recorder and on exit, sends the recorded file to the s3 bucket. Admission controller is used to deny access to k8s pods using kubectl exec. Also Admission controller is used to deny use of any other ephemeral containers apart from the ones provided.
Can it be deployed onprem? * Yes, all the components are chosen such that it can be deployed onprem in air-gapped environments too.
Why did I work on it? * My aim with this exercise was to actually build a realistic tool that can help real users... There are a million other case studies where you could find AI helping users with one-off tasks. One never gets things right in the first iteration when building anything at scale. I assumed there would be rework, debugging, deprecating existing modules, and designing different modules. For scale and multi-tenancy, I changed the storage model multiple times. I wanted to check if AI would actually be effective in all these phases and what kind of issues I would face... working on an article explaining the challenges, cost, best practices for developing an AI-assisted app, and more importantly, what security practices could be followed while doing so. Note: I ran security scans on the auto-generated code and manually verified the findings. Then I used Claude itself to fix those security findings. To summarize, I believe AI could independently orchestrate full-blown enterprise apps.