Cobsidian

English · 简体中文

Turn AI conversations into linked Obsidian knowledge, safely.

Cobsidian is an agent-agnostic workflow skill for maintaining an Obsidian or Markdown vault. It searches before creating, previews the proposed change, asks for one plain-language confirmation, writes atomically, and validates the result. Integrity hashes stay in the execution layer; users do not copy confirmation codes.

It is not a hosted service or an Obsidian plugin. Your coding agent runs the workflow against a local folder of Markdown files.

Quick Start · How It Works · Install · DeepSeek Harness · MCP Server · Compatibility · Latest Release

Use the agent you already know: Claude Code · Codex CLI · GitHub Copilot CLI · Kimi Code · OpenCode · Pi · Antigravity

<p align="center">_{Synthetic demo vault. No private notes, paths, or credentials.}</p>

Quick Start

Run a read-only dry run against the bundled demo vault:

git clone https://github.com/Totoro-qaq/Cobsidian.git
cd Cobsidian
python skills/cobsidian/scripts/dry_run.py examples/demo-vault --topic "AI Conversations" --mode learning --text "agent workflow notes" --json

Then point your agent at skills/cobsidian/SKILL.md:

Use Cobsidian to organize this material into my Obsidian vault.
Vault: /absolute/path/to/obsidian-vault
Run a dry run first, check duplicates, suggest backlinks, and wait for confirmation before writing.

What Cobsidian Does

Stage	What Cobsidian makes visible
Read	Resolves the vault and builds identities from filenames, H1s, titles, and aliases.
Decide	Reports `create
Review	Returns duplicate risk, backlink suggestions, and an exact patch plan.
Approve	Asks Apply this reviewed change?; the selected plan ID stays internal.
Write	Verifies the internal plan and file hashes, writes atomically, validates, and keeps rollback available.

The result is durable Markdown with useful [[wiki links]], not a second copy of a note that already exists.

Before / After

flowchart LR
    A["AI chat, logs, project notes"] --> B["Cobsidian dry run"]
    B --> C["Search existing vault notes"]
    C --> D{"Machine action<br/>create | append | blocked"}
    D --> E["Note plan<br/>single-note | multi-note | report-only<br/>split = multi-note"]
    E --> F["Patch preview + simple approval"]
    F --> G["Atomic write + validation"]
    G --> H["Rollbackable transaction"]

Before	After
Useful answers disappear in chat history	Reusable notes stay in the vault
Repeated prompts create near-duplicates	Existing identities are checked first
Links are guessed while writing	Backlinks come from actual vault notes
Agent edits are hard to audit	A reviewed, integrity-bound plan precedes every write

How It Works

…