Obsidian PKB Guide: Build a Personal Knowledge Base for AI Agents

A personal knowledge base stalls when you only bolt a chat panel onto a notes app. The practical design is different: treat your Obsidian vault’s local Markdown as shared memory, and run CLI agents with the vault as their working directory (cwd). That gives you a loop you can stage—capture, organize, and distill—without pretending one tool owns everything.

📑Table of Contents

1. Why the vault should be the agent workplace
2. The 2026 PKB map — human thinking space vs agent-maintained knowledge
3. Level 0 — make the vault the agent cwd with Claudian
4. Role split — Hermes capture, CLI distill, and alternatives
5. Comparison table — Claudian / Hermes / second-brain / bare CLI
6. Staged checklist — your next actions
7. FAQ
8. Takeaways — the decisions to keep

What this article covers

This article explains that staged design. Product claims are checked against the Claudian repository and Hermes Agent docs, plus independent guides and operator write-ups.

Discovery popularity alone is not enough to decide adoption.

Learning outcomes

---

Why the vault should be the agent workplace

If knowledge lives only in chat transcripts, reuse collapses when the context window resets. If Markdown in the vault is the source of truth, humans and agents share the same durable surface: read, write, search, and bash against one tree.

Local sovereignty stays with you

Independent analysis frames Obsidian as a private, malleable, independent thinking space built on local Markdown (note.com / kagawatomo).

The agent’s unit of work becomes clear

Claudian documents embedding coding agents so the vault is the working directory, with sidebar multi-step work rather than paste-only chat (GitHub: YishenTu/claudian).

Roles can be split

External capture and in-vault distillation can live on different layers. Hermes is documented as an autonomous agent with tools, MCP, messaging gateways, and persistent memory—not tied to a single IDE (Hermes Agent Docs).

Boundary to keep

What you should leave with is not “the best stack,” but a boundary: what humans still decide, and what agents may write.

---

The 2026 PKB map — human thinking space vs agent-maintained knowledge

A realistic 2026 framing is division of labor, not full automation.

Independent framing of the split

The same independent note maps Obsidian’s role as a human thinking space, then places patterns such as Karpathy’s LLM Wiki (raw/wiki/schema plus ingest/query/lint), GBrain, and Hermes Agent under “agents maintain knowledge.” The open question is how far you should delegate (note.com / kagawatomo).

What Hermes docs add

Hermes docs describe skills that improve over time, persistent memory, a large tool set, MCP, and messaging gateways.

Official user stories collect community workflows that write daily logs, tasks, and long-lived vault Markdown—evidence that Obsidian-backed agent memory is practiced beyond a single tutorial.

Three areas to split before automation

Split these three areas before you automate:

Area	Keep with humans	Delegate to agents
Judgment	Publish/no-go, priority, final approval	Drafts, summary candidates, proposed diffs
Maintenance	Folder conventions and policy	Ingest, lint, link repair, daily note generation
Connectivity	Which external sources are trusted	Fetch/write-back only through allowed tools

Why judgment boundaries come first

If judgment boundaries stay vague, overwrite incidents and sensitive-file leakage become more likely—even when the product features look strong.

Sources (public descriptions as of July 2026):

• note.com (kagawatomo)
• Hermes Agent Docs
• Hermes User Stories

---

Level 0 — make the vault the agent cwd with Claudian

Level 0 means: keep the vault open and run a CLI agent from the sidebar. The core is not a decorative chat pane; it is agent execution whose cwd is the vault.

What the product baseline claims

The Claudian repository describes embedding Claude Code, Codex, Grok, Opencode, Pi, and similar agents with multi-tab chat, inline edit plus diff, slash Skills, @mentions, Plan Mode, and MCP.

Requirements include:

• Desktop Obsidian 1.7.2+
• Claude Code CLI required (Codex and others optional)
• Install via Community Plugins or release files (main.js, manifest.json, styles.css)

Independent checks to keep side by side

• Sébastien Dubois’s independent summary restates the same capability set and Desktop/1.7.2+ bar.
• NateCue’s guide adds operational detail: connect a preinstalled CLI, use sidebar chat and inline diffs, and treat MCP as a hub.

Privacy and CLI-path notes

• The same guide notes that the plugin itself does not add telemetry (data goes to the configured provider), and covers CLI path troubleshooting under nvm.

Source

An operations pattern from a six-month write-up

Qiita (tasekino) reports roughly six months of Claude Code inside Obsidian via Claudian:

• Role-bearing folders such as daily / notes / news / memo
• Pass the open note as context
• One custom slash (for example /daily) before a catalog of automations
• After overwrite incidents, accumulate must-confirm rules in CLAUDE.md (fail → note → rule)

Level 0 guardrails to install first

• Default to Plan Mode and diff preview
• Ban bulk overwrite in project instructions
• Define ignore/allow lists for sensitive paths before wide agent access
• Do not force Level 0 on mobile-sync-only setups (Desktop is assumed)

Sources (July 2026):

• Claudian on GitHub
• dsebastien
• NateCue
• Qiita (tasekino)

---

Role split — Hermes capture, CLI distill, and alternatives

Loading capture and distill into one agent blurs auth and failure boundaries.

A workable three-way split

1. Hermes for capture — external messages, tool credentials, gateway intake (tools/MCP/messaging in the official docs).
2. CLI agents via Claudian for distill — organize, summarize, lint, and repair inside the vault.
3. Humans for judgment — priority, publishability, and whether delegation is actually safe.

If you already run Hermes

If you already run Hermes, keep external auth there and limit vault mutation to Claudian + CLI. If you do not, start at Level 0 (Claudian only) and add capture later.

Alternative architecture

As an alternative architecture, obsidian-second-brain publishes persistent Markdown memory for Claude Code/Codex-style agents, many vault-maintenance commands, hybrid search, and scheduled agents.

That shows “CLI agent + vault second brain” is a broader ecosystem pattern, not a single template.

Decision branches

• Edit while notes stay open → Claudian Level 0
• Heavy external inflow (chat/mail/tools) → consider Hermes capture
• Want thick skill packs / scheduled maintenance → compare second-brain-style repos
• Mobile-first or no CLI → outside this article’s Level 0 premise

Sources:

• Hermes Agent Docs
• obsidian-second-brain
• Operational framing: note and Qiita

---

Comparison table — Claudian / Hermes / second-brain / bare CLI

Pick by which surface you want the product to own—not by a generic “best tool” ranking.

Surface and cwd

Dimension	Claudian + CLI	Hermes Agent	obsidian-second-brain	Bare CLI only
Primary surface	Obsidian sidebar / inline	CLI, Desktop, messaging gateway	Vault skill packs + CLI	Terminal
Vault as cwd	Core product behavior	Operational wiring (Obsidian in user stories)	Persistent-memory premise	Manual cd

Strengths, prerequisites, and stage

Dimension	Claudian + CLI	Hermes Agent	obsidian-second-brain	Bare CLI only
Strengths	Note context, diffs, Plan Mode, MCP	External connectivity, self-improving skills, memory	Many maintenance commands, hybrid search, scheduled agents	Minimal dependencies
Prerequisites	Desktop Obsidian + Claude Code CLI, etc.	Install + provider setup	Repo / Skills install	CLI + path conventions
Best stage	Level 0 (work inside)	Level 1+ (bring work in from outside)	Alternative / extension architecture	Experiment / minimal PoC

How to read the table

Sources (July 2026):

• Claudian
• dsebastien
• NateCue
• Hermes Docs
• Hermes User Stories
• obsidian-second-brain

An “orchestra” layout is one role-split example. Verify each component against public docs, then stage to your capture sources and Desktop constraints.

---

Staged checklist — your next actions

Do not “install everything.” Name today’s level and the judgments humans keep.

Level 0 (today–this week)

• ✓ Desktop Obsidian 1.7.2+ and Claude Code CLI (plus Codex CLI if needed) work
• ✓ Install Claudian and confirm the vault is the agent cwd
• ✓ Default Plan Mode/diffs; ban bulk overwrite in CLAUDE.md (or equivalent)
• ✓ Write ignore/allow lists for sensitive folders first

Level 0.5 (once the loop is running)

• ✓ Add one folder convention pair (daily / notes) and one slash command (for example /daily)
• ✓ Weekly fail → note → rule review

Level 1 (move capture outside)

• ✓ Evaluate isolating external-tool auth outside the vault (Hermes or similar)
• ✓ Tabulate inbox → organize → distill ownership for humans vs agents

Go / no-go judgment

Situation	Judgment
You want to edit while notes stay open	Claudian Level 0 fits well
External inflow is high	Consider Hermes capture
You want non-Claudian skill packs	Compare second-brain-style options
Mobile-first / no CLI	Outside this Level 0 premise—do not force it

⚠️ Bulk rewrite without Plan Mode/diffs, unrestricted access to sensitive notes, and forcing Level 0 on mobile-only sync are high-risk patterns. Write permission boundaries before broad agent access.

---

FAQ

Q1. Is Claudian just a chat UI embedded in Obsidian?

A. No. Official and independent guides describe a CLI agent whose cwd is the vault—with read/write/search/bash, Plan Mode, MCP, and related controls—not a paste-only conversation pane.

Q2. Is Hermes required?

A. No. Level 0 can stand on Claudian + CLI. Hermes is a Level 1+ candidate when you want external capture, credentials, and gateways separated.

Q3. Can phone-only Obsidian do the same thing?

A. Claudian assumes Desktop. Mobile sync alone breaks the Level 0 premise; design another path.

Q4. Is one practical template article enough?

A. It is a useful discovery and setup entry. Product requirements, privacy notes, and alternatives still need official README plus independent guides (dsebastien, NateCue, Qiita, note, second-brain). This article is for that cross-check.

Q5. Should the agent own the whole vault?

A. Safer default: keep human judgment zones. Install CLAUDE.md guardrails, Plan Mode, and sensitive-path exclusions first—the same pattern as operator reports that turned overwrite incidents into rules.

---

Takeaways — the decisions to keep

• Designing PKB as “vault = shared memory + agent cwd” makes capture and distill operable as a staged loop.
• Level 0 is a minimal setup you can verify from Claudian’s public requirements and independent guides; multi-month ops notes (for example on Qiita) add failure-mode detail.
• Hermes and second-brain are comparison axes for capture separation and alternative skill packs—not mandatory installs.
• Your next action: confirm Desktop/CLI feasibility, advance one step on the Level 0 checklist, and write the judgments humans still own.

Decide adoption from public docs, independent verification, and your constraints—not from discovery signals alone.