Most “AI for notes” products sell a plugin: install, connect an API key, chat inside the vault. That works—until you want the same agent that already refactors your codebase to treat your notes as a first-class project.
A simpler frame is spreading: the vault is a folder. A coding agent (Codex, Claude Code, and peers) is a filesystem worker. “Connecting” is cd into the vault and running the agent—not wiring a proprietary bridge. The discipline file (AGENTS.md / CLAUDE.md) is the schema that keeps the agent useful instead of chaotic.
This Discover piece explains that repo-as-notes pattern, how official agent docs make a root instruction file work, where to put guardrails, and how it differs from an LLM Wiki that compiles pages (Discover #1) or a local semantic finder (Discover #3).
Verification status: Not tested by Cyber-Taoist editorial. Claims follow the cited practitioner write-up and official agent docs. Treat this as an explainer of a method worth trying, not a lab report.
The insight: connection = folder + agent
Practitioner view (Aishwarya Srinivasan): An Obsidian vault is ordinary files on disk. Codex (or a similar coding agent) already reads and writes directories. The setup is: install Obsidian, keep a light folder layout (notes/, daily/, projects/, references/), drop an AGENTS.md at the vault root, then run the agent from that directory. Typical jobs include summarizing across notes, spotting contradictions, and turning a week of meeting notes into a project brief—without asking the agent to rewrite .obsidian config.
Editorial analysis: This reframes “AI second brain” from chat widget inside the note app to agent with a project brief for your knowledge repo. It fits people who already trust Git, diffs, and coding agents. It is a poor default if you only want a one-click summarize button and never open a terminal.
Why AGENTS.md is the schema
Author / product view (OpenAI Codex docs): Codex reads AGENTS.md before acting. Discovery is layered: a global home config plus project files from the repo root down to the working directory. Same-directory overrides win; deeper files override shallower ones; merged instruction size is capped (default on the order of tens of kilobytes). You can point fallback filenames at existing team guides so the vault does not need a second parallel bible.
Author / product view (Anthropic Claude Code docs): CLAUDE.md is human-written persistent instruction; the tool can also honor AGENTS.md. Prefer short, checkable rules (paths, naming, commands)—on the order of under ~200 lines—and push long procedures into skills or path-scoped rules. Local preferences can live in a gitignored local file; shared repo rules stay in the committed instruction files.
Editorial analysis: For a single vault serving more than one agent, put shared rules in AGENTS.md and let Claude-side files import or mirror them. The file is not magic—it is a contract: what the agent may touch, how notes are named, and what “done” looks like for a session.
A practical vault layout and deny list
Practitioner view (Dexio guide, Claude Code + Obsidian): Start in the vault with the agent CLI. A workable layout separates inbox/, raw/ (immutable sources), notes/, plus navigation aids like index.md and a session log.md. A vault instruction template typically asks for wikilinks, search-before-create, and a hard ban on editing .obsidian. When Codex and Claude Code share one vault: keep shared policy in AGENTS.md, import it from CLAUDE.md, deny edits to raw sources and Obsidian config in agent settings, and review every session with git diff.
Editorial analysis: Deny paths and Git review are the difference between “helpful intern” and “silent vandal.” Raw captures and app config should be read-only to the agent. Treat plans as plans—not as evidence that work finished (a principle OpenAI’s Codex workflow cookbook stresses for software harnesses; it transfers cleanly to note ops).
Try path (minimal)
- Put the vault under Git (even if you never push).
- Add a root
AGENTS.md: folder meanings, naming, “never edit.obsidianorraw/,” prefer edit-over-create, end with a short session log entry. - From the vault root, start Codex (or Claude Code) and ask for a read-only pass: contradictions on one topic, or a brief from last week’s meeting notes.
- Inspect
git diff; keep, revise, or revert before a write-heavy follow-up. - Only then allow targeted edits inside
notes/—still never bulk-rewrite history.
Editorial analysis: First sessions should prove judgment, not throughput. If the diff feels wrong, tighten the schema before unlocking writes.
How this sits next to LLM Wiki and plugins
| Job | Lean toward |
|---|---|
| Agent reads/writes the live vault with human-gated diffs | Codex / Claude Code + AGENTS.md (this piece) |
| Model maintains a compounding compiled wiki between you and raw sources | LLM Wiki pattern (Discover #1) |
| “Notes like this one” while you draft, often local | Smart Connections (Discover #3) |
| Human Distill layers that make notes findable | Progressive Summarization (Discover #2) |
Editorial analysis: Repo-as-notes is an operations surface. LLM Wiki is a compile surface. You can use both: raw + notes under Git for the agent; a separate wiki tree the model maintains. Do not confuse “the agent edited three files” with “the wiki is consistent.”
Limits and failure modes
- Cloud coding agents send vault context to the model provider—unlike local-only embedding discovery. If that is unacceptable, use a local model stack (queued Discover) or keep secrets out of the vault.
- Official docs describe general repos; vault-specific rules (wikilinks,
.obsidian, evergreen titles) are yours to write. - Paid practitioner essays may truncate steps behind a paywall; verify against official
AGENTS.md/CLAUDE.mdpages. - Without Git + deny rules, an eager agent will “helpfully” reorganize your life.
- This article did not run Codex against a real vault (
verification: none).
Bottom line
You do not need a proprietary “brain plugin” to get an agent into your notes. Treat the vault as a folder, the agent as a filesystem worker, and AGENTS.md as the schema. Keep raw and app config off the write path, review every diff, and reserve compiled-wiki dreams for a deliberate LLM Wiki layer. That is how coding-agent habits transfer to knowledge work without pretending the model owns your judgment.
Sources
- How to Set Up Your Second Brain with Obsidian and Codex — Aishwarya Srinivasan (2026-07-17)
- Custom instructions with AGENTS.md — OpenAI / ChatGPT Learn
- CLAUDE.md / memory — Anthropic Claude Code Docs
- Obsidian + Claude Code guide — Dexio (updated 2026-10-01); commercial product at end of page
- Iterating development workflows with Codex — OpenAI Cookbook (harness / human-gate principles)