Claudian
An Obsidian plugin that makes your vault the working directory for Claude Code, Codex, and other coding agents.
- Source repo
- YishenTu/claudian
- Stars
- ★ 16k
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 49/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Researchers drafting long documents in Obsidian who want an agent to search, cite, and rewrite across many notes in the vault.
- Not a fit if
- People who need this on Obsidian mobile, which is unsupported
- Users who want lightweight note-taking without installing and maintaining external CLIs
- Teams whose vault content must never leave the machine to Anthropic, OpenAI, or xAI
- Source review
- 49/100 · Major gaps
What does this agent do, and when should you use it?
Claudian is an Obsidian plugin that embeds AI coding agents — Claude Code, Codex, Grok Build, OpenCode, Pi, DeepSeek Harness, and more — inside your vault, so the vault itself becomes the agent's working directory with file read/write, search, bash, and multi-step workflows available out of the box. It runs as a sidebar chat with tabs, single- or dual-pane session management, a Zen mode compact composer, inline editing with word-level diff preview, slash commands and Skills, @mentions for vault files and other sessions, and cross-session messaging. Models are not hard-coded; the plugin discovers the catalog reported by each configured provider runtime and you enable what you want. Per-provider adapters (Claude Agent SDK, Codex app-server, Grok ACP, OpenCode ACP/HTTP, Pi RPC, DeepSeek Remote) share ACP mechanics and the agent's native MCP configuration. It ships under MIT, is installable from Obsidian Community Plugins, is desktop-only on Obsidian v1.13.0+, and requires at least one harness CLI plus a subscription or API provider you supply.
The sidebar chat is the main surface: you give an instruction and the agent reads, writes, edits, and searches files inside the vault and runs commands there. Inline Edit takes the selected text or the cursor position plus a hotkey and returns a provider-backed rewrite shown as a word-level diff before you accept it. Typing / or $ opens reusable prompt templates and Skills from user- and vault-level scopes; typing @ references vault files, folders, and other Claudian sessions. Cross-session messaging lets one agent read, message, steer, or interrupt other Claudian sessions or start new ones on any provider, and /side or /btw opens a temporary side conversation without touching the main chat. External tools attach through each coding agent's own CLI-managed MCP configuration, and a 127.0.0.1-only listener serves Claudian's session tools to OpenCode, DeepSeek, and Grok sessions.
- Researchers drafting long documents in Obsidian who want an agent to search, cite, and rewrite across many notes in the vault.
- Personal knowledge-base owners who want to reorganize, rename, and complete Markdown notes and attachments through natural-language instructions.
- Developers already subscribed to Claude Code or Codex who want the same coding-agent workflow applied to writing and reference material.
- Users comparing output across providers (Claude, Codex, Grok, OpenCode, and others) from one interface.
- Editors doing small wording passes, relying on inline edit and the word-level diff to confirm exactly what changed.
- Advanced users who want agents working in background sessions and coordinating through cross-session messages.
How do you install or deploy this agent?
Install from Obsidian Community Plugins: open Obsidian → Settings → Community plugins → Browse, search for "Claudian", install, and enable it. You can also install directly from the community plugin page.
Install from source (development): clone the repository into your vault's plugins folder, then install dependencies and build.
cd /path/to/vault/.obsidian/plugins
git clone https://github.com/YishenTu/claudian.git
cd claudiannpm install
npm run buildThen enable "Claudian" under Settings → Community plugins. Prerequisites: Obsidian v1.13.0+ (desktop only), at least one harness CLI (Claude Code, Codex, DeepSeek Harness, Grok Build, OpenCode, or Pi), and a compatible subscription or API provider.
How do you use this agent?
Open the Claudian interface from the ribbon icon or the command palette and talk to the agent; it reads, writes, edits, and searches inside the vault and runs commands there.
Key entry points:
/ slash command templates
$ Skills
@ reference vault files, folders, or other Claudian sessions
/side /btw temporary side conversationIf a provider CLI is not auto-detected, expand the installation card under Settings → Claudian → Providers → Claude Code and set the CLI path. Find the executable with which claude on macOS/Linux or where.exe claude on Windows. If an npm-installed CLI reports Node.js missing, add its bin directory under Settings → Claudian → General → Environment → Shared environment:
PATH=/path/to/node/binIf your terminal subscription works but the plugin reports authentication_failed, clear the inherited credential under Settings → Claudian → Providers → Claude Code → Environment → Custom variables:
ANTHROPIC_API_KEY=What are this agent's strengths and limitations?
- Provider model catalogs are discovered at runtime (Discover / Browse models) rather than hard-coded, so new models do not require editing the installed bundle.
- One plugin covers six harnesses — Claude Code, Codex, DeepSeek Harness, Grok Build, OpenCode, and Pi — switchable in provider settings.
- Concrete writing-oriented interactions: inline edit with word-level diff, Zen mode composer, @mentions, and cross-session messaging.
- Privacy boundaries are stated explicitly: no telemetry beacons, network activity limited to provider calls, configured MCP endpoints, and a 127.0.0.1-only listener for session tools.
- MIT-licensed and available directly from Obsidian Community Plugins, so no custom packaging is needed.
- Desktop only (macOS, Linux, Windows); mobile Obsidian is not supported.
- Requires you to install at least one external harness CLI and bring your own subscription or API — the plugin ships no model of its own.
- Obsidian launched from the Dock on macOS/Linux does not inherit the terminal PATH; resolution relies on the login shell or a manually configured CLI path, which trips up version-manager setups.
- On Windows,
.cmdand.ps1wrappers are discouraged in favor ofclaude.exeorcli-wrapper.cjs, adding configuration friction. - Pull requests adding new providers are explicitly declined, and Collab has moved to a separate plugin, so the surface is bounded by the maintainer's scope.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Claudian This agent | 49 · Major gaps | IDE extensionFree + model costs | ★ 16k | today | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| IWE Knowledge Graph & Agent Memory | 73 · Some gaps | CLIFree | ★ 1.8k | 2d ago | Rust | Codex · Claude Code · Claude.ai |
| Cindy | 52 · Major gaps | Desktop appFreemium | ★ 3.2k | today | TypeScript | Codex · Claude Code |
| Memmy — Personal AI Agent & Local Memory Hub | 42 · Major gaps | Desktop appFreemium | ★ 2.1k | today | TypeScript | Codex · Claude Code · OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
The README privacy section explicitly lists data sent to APIs (input, attachments, images, tool outputs) and destinations (Anthropic/OpenAI/xAI etc.), and states no telemetry with network activity limited to explicit provider calls, configured MCP endpoints, and a 127.0.0.1 listener — data_flow_transparency earns 2. However the plugin makes the whole vault the agent working directory with file read/write, bash, and multi-step workflows; the README does not state default permission modes, whether destructive actions (delete/overwrite/command execution) require confirmation, or any rollback/undo mechanism, so least_privilege, user_confirmation, and rollback score 1. Sensitive data handling only advises sharing variable names rather than secret values; no key storage/encryption guidance, score 1. Dependency security has npm audit, dependency-review, lockfile checks, and pinned overrides, score 2, though many external CLIs/SDKs lack a vulnerability inventory. External effects (bash, cross-session messaging, MCP) lack boundary constraints, score 1. Source attribution is only README/LICENSE naming Yishen Tu with unverified publisher identity, score 1.
Clear layered structure (core/providers/features) and CI with typecheck, lint, architecture-boundary tests, cross-platform smoke, and artifact validation support self_consistency at 2. Dependency availability requires Node >=24 <25, Obsidian v1.13.0+, desktop-only, and external CLIs (Claude Code/Codex etc.) with PATH probing, a high bar, score 1. Failure messages: the Troubleshooting section covers ENOENT, missing Node, auth conflicts with concrete recovery steps, score 2.
Audience and scenarios are clear (Obsidian users running coding agents in their vault), score 2. Capability boundaries are stated (desktop-only, provider list, no new-provider PRs), score 2. Trigger precision: slash commands, @mention, cross-session messaging are only summarized with no conflict/ambiguity handling, score 1. Environment fit covers macOS/Linux/Windows PATH and CLI detection differences, score 2.
Information architecture has a directory tree and feature list, score 2. Install notes cover community plugin and source builds, score 2. Naming stability: version 2.3.26 with OpenCode v1 support ending 2026-10-30 signals breaking-change risk, score 1. Examples and FAQ limited to Troubleshooting with no usage examples or FAQ set, score 1. Known limitations only scattered (desktop-only, provider compatibility), not consolidated, score 1. License is full MIT text, score 3. Versioning/changelog has only a version number and preversion script, no CHANGELOG, score 1. Maintenance responsibility has contribution guide and contact, but publisher identity is unverified, score 1.
Output usability offers inline-edit diff preview, Zen mode, and tabs for real writing workflows, score 2. Marginal value of embedding coding agents in a note vault is differentiated, score 2. Cost-benefit requires subscription/API fees plus external CLI and Node 24 setup, with no cost discussion in README, score 1.
Feature claims in README lack corresponding code or test references, claim_traceability scores 1. Test mocks and harness plus CI partially corroborate engineering practice but cannot cross-verify README feature claims, score 1. README mixes facts and inferences (e.g., 'no telemetry') without evidence sourcing, fact_inference_separation scores 1.
- The plugin uses the entire vault as its working directory with bash and file-write capability; the README does not state default permission modes or confirmation/rollback for destructive actions — restrict permissions and back up before use.
- Publisher identity is unverified; verify the Obsidian community plugin source and signature before installing.
- Depends on external CLIs and Node 24, a high setup bar; PATH detection failures can break functionality.
- OpenCode v1 support ends 2026-10-30, posing a breaking-upgrade risk.
- Claims such as 'no telemetry' lack code-level evidence and cannot be verified by static review.
FAQ
Does the plugin itself cost anything?
Where does my vault content go?
Why can't the plugin find my agent CLI?
spawn claude ENOENT and Claude Code CLI not found. Obsidian does not inherit the terminal PATH on macOS/Linux, so Claudian reads it once from your login shell; you can also set the CLI path manually in provider settings.