Published: 2026-08-01

Hermes + Obsidian + Ollama: a fully local, free agent notes setup

Chapters / key moments (click to jump — plays here on the page)

Fahd Mirza joins three pieces — Hermes agent, an Obsidian vault, and a local model served by Ollama — into a note-taking assistant that runs entirely on his own Ubuntu machine with no API key and no cloud calls. The agent creates notes, searches the vault, reads results back, and links notes together using Obsidian's wiki-link syntax. In the last test it reaches for the Obsidian skill unprompted, without being told the skill name or the vault path.

Source video

"Hermes-Agent + Obsidian + Ollama: Your Notes, Now Hands-Free" by Fahd MirzaWatch on YouTube →

Key Takeaways

  • The whole stack is local and free. Hermes supplies the agent loop, Ollama serves the model on localhost, and Obsidian is the storage layer. Nothing leaves the machine and no paid API is involved — the demo runs on a single-GPU Ubuntu server.
  • An Obsidian vault is just a folder of markdown files. There is no database and no proprietary format, which is why an agent with ordinary file tools can work with it. Hermes ships a bundled Obsidian skill that reads, searches and writes those files, so nothing custom has to be built.
  • Use a dedicated Hermes profile. Mirza creates a separate profile for this setup so the local-model configuration stays isolated from his other model profiles rather than overwriting them.
  • Ollama needs no real API key, but the field must not be empty. Hermes expects an API key value even when the backend is local, so a placeholder string is enough to get past the check. Config is verified afterwards with the Obsidian skill's own config-show command, which prints the vault path and model settings.
  • Wiki-links let the agent build the graph. Asked for a second note linked back to the first, the agent wrote Obsidian's double-bracket link syntax itself — that is what turns a flat folder of notes into a connected graph.
  • Skill selection worked without being named. The final test used a neutral prompt — write a short note on a given topic — with no mention of Obsidian and no vault path. The agent identified it as an Obsidian job, loaded the skill, and resolved the path from its configuration on its own.

Gotchas

  • A smaller local model is less efficient, not less correct. On the linked-note task the agent wrote the file twice before recognising what it had done. The end result was right, but Mirza notes a stronger model would have taken fewer steps — worth budgeting for if you run a small quantised model.
  • Point the skill at the vault explicitly. The vault path is set as its own configuration value after the profile is created. If it is missing, the skill has nothing to read or write.

Weekly Digest — In Your Inbox

Get the week's top AI agent news, updates, and guides — every Friday.