← sources

I built a self-hosted platform where Ollama agents work on real servers, fully offline, and ask before running anything

source: reddit (reddit:ollama)published 2026-10-28 10:13ingested 2026-10-29 02:42

retained text

    I wanted AI agents that could actually help in an IT office, without sending our infrastructure to the cloud and without blindly trusting them with root. So we built **Orlok**, a self-hosted platform of *virtual colleagues* that runs **fully offline with Ollama**.
    
    **What it looks like (first screenshot)**
    
    "Nobody can print since this morning, can you check app-01?" → the agent runs a read-only check over PowerShell, finds the spooler stopped, recognizes the issue from the office wiki, and **waits for approval** before running `Restart-Service Spooler` (allow once / always allow / deny).
    
    **The local-model side**
    
    - **Ollama as a first-class provider**: no cloud account needed; Anthropic, OpenAI and compatible APIs are there too if you want them
    - **Profiles = model + mode** (second screenshot): e.g. a local 32B model that may only *ask* to run commands, or one that can't run commands at all and just answers from the wiki
    - **Bundled embedding model**: the wiki is indexed locally, so your pages never leave your network to become searchable
    - Wiki, tools and guardrails live in Orlok, not in the model: switching from cloud to local changes neither the rules nor the knowledge
    - Every call is logged; local models are counted but not priced, cloud ones with their cost and monthly limits
    
    **What the agents can do**
    
    - Read and update an office **wiki**: import from Obsidian, Notion or Joplin (passwords and keys stripped), PDFs and scans with local OCR, nightly review for contradictions and stale pages
    - Work on **Linux over SSH** and **Windows over WinRM/OpenSSH**, with multi-step plans you approve as a whole or step by step
    - Commands are **parsed, not regex-matched**; each user's own encrypted credentials are used, from an isolated runner container; append-only audit log
    - Web, installable phone app and Telegram
    
    **Install:** `git clone` + `docker compose up`, then a first-run wizard where you add Ollama as a provider.
    
    **License, upfront:** source-available (BSL 1.1), free for personal and non-commercial use and for evaluation; each version becomes Apache 2.0 after 4 years. Production use for a company needs a commercial license.
    
    I'm curious which local models you'd trust with tool calling on real servers. Feedback very welcome!
    
    Website: https://orlok.xyz
    Code: https://github.com/rylos/orlok
    Docs: https://orlok.xyz/docs

document fetched 2026-10-29 00:43 · canonical https://reddit.com/r/ollama/comments/1xhk3vy/i_built_a_selfhosted_platform_where_ollama_agents