---
id: "lib-2026-10-08-personal-os-migration-to-server-side-rag-with"
type: "library"
title: "Personal OS migration to server-side RAG with sandbox"
title_original: "Я мигрировал Personal OS с git repo на server-side RAG с песочницей"
lang: "RU"
org: "LLM под капотом (Telegram, Rinat Abdullin)"
source_type: "Practitioner post"
date: "2026-10-08"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-08-personal-os-migration-to-server-side-rag-with/"
original_url: "https://t.me/llm_under_hood/957"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Practitioner post from the Russian-language Telegram channel LLM pod kapotom (@llm_under_hood) reporting a concrete architecture change with measured task times; useful practice, but one user's setup and two timings."
summary_basis: "full-text"
original_minutes: "4"
tags: ["ai technology", "enablement", "agent knowledge base", "llm wiki"]
frames: ["tacit-to-explicit"]
---

# Personal OS migration to server-side RAG with sandbox

Original title: Я мигрировал Personal OS с git repo на server-side RAG с песочницей

Tier B (Useful, limited): Practitioner post from the Russian-language Telegram channel LLM pod kapotom (@llm_under_hood) reporting a concrete architecture change with measured task times; useful practice, but one user's setup and two timings.

Summary based on the full text.

## In short

A post (edited 8 October 2026) in the Telegram channel LLM pod kapotom, "Experiment 12" in the author's series on a "Personal OS": a personal knowledge base worked on by AI agents. About 1,500 documents moved from git plus Obsidian to a self-written versioned file server. Agents found documents faster but took longer to edit them. The fix was to drop most tools and let the agent send small Python scripts to run in a sandbox on the server, which cut a typical task from 6.5 to 3.5 minutes.

## Key numbers

- ~1,500: documents in the Personal OS knowledge base moved from a git repo to a server
- 6.5 -> 3.5 min: time for the agent to file a conversation note before and after switching to server-side Python
- 2: tools left in the MCP API and CLI: start and run

## What it reports

- Earlier stages of the experiment: git plus Obsidian (8 February 2026), Personal OS structure (27 February), comparison with Karpathy's LLM Wiki (5 April), an agent persona (23 May).
- New server: Go, SQLite, content-addressed storage, an MCP API, a CLI installed for Codex, and Borg Backup to a remote server.
- Gains: no git push/pull juggling; versioning is global, so an agent change touching a thousand files across workspaces can be rolled back at once; search works the same for every user; built for several people and their agents working on shared material.
- Cost: editing remote files was slow. Filing a conversation summary into the right workspace (including reworking a 70 KB file) took the Sol 6.1 model 6.5 minutes.
- After the change: the agent writes a short Python script that runs in an isolated per-session sandbox on the server. The same task now takes 3.5 minutes, about half on retrieval and half on writing.
- The MCP API and CLI shrank to two commands: start (open a session/commit and receive the SDK description) and run (execute Python on the server).
- Advice: if an AI-native system grows a complex set of APIs and tools, give agents Python instead; sandboxing options range from jails to Firecracker VMs.

## How they got it

Personal engineering experiment on the author's own system. Timings are single observations of one representative request, before and after the change.

## Limits

- One user, one model, one task type; no repeated measurements.
- Running agent-written code server-side needs strict isolation; security is mentioned only through sandbox options.

## Why it matters for companies

Many firms are building shared knowledge bases for staff and agents. This post shows two design points: version the whole store so an agent's bulk edit can be undone, and keep the agent interface small. A sandboxed code runner can replace dozens of narrow tools and be faster. The tradeoff is that the security and audit work moves to the sandbox.

## Key passages in translation

> "Можно откатить изменение агента, даже если он поменял тысячу файлов в разных workspaces одновременно"

"You can roll back an agent's change even if it modified a thousand files in different workspaces at the same time."

> "агенты находят документы быстрее, но стали дольше копошиться с их изменением"

"Agents find documents faster, but now take longer fiddling with changing them."

> "Самое забавное, что этот интерфейс оказался питоном."

"The funniest part is that this interface turned out to be Python."

> "агент пишет маленький скрипт, который отправляется на сервер и в изолированной песочнице работает с изменениями в рамках сессии там"

"The agent writes a small script that is sent to the server and works on the changes there, in an isolated sandbox, within the session."

> "если у вас у какой-то AI Native системы получается сложный интерфейс с кучей API-шек и инструментов - попробуйте просто дать агентам Python"

"If some AI-native system of yours is ending up with a complex interface full of APIs and tools, try simply giving the agents Python."

Original: https://t.me/llm_under_hood/957