Beyond Chat: Building an AI-Native Knowledge System with Claude Code and Obsidian

On This Page

    I got tired of copying context.

    Every new chat session, I’d spend the first few messages re-explaining who I am, what I’m working on, what I already tried. Years of notes sitting in Obsidian. Months of useful conversations with Claude. None of it connected. The AI knew nothing about my vault, and my vault knew nothing about the AI services I use.

    So I built the connection myself.

    The Limits I Kept Hitting

    I’ve used ChatGPT, Gemini, Claude - web apps, desktop apps, API integrations. They all hit walls:

    • Web interfaces lose context constantly. New session, new brain.
    • Desktop apps are better but still siloed. They can’t touch my files without clunky workarounds.
    • Claude Code changed things - finally, an AI that could operate on my filesystem. But the CLI is optimized for development, not knowledge management.

    I tried building multi-agent systems myself. Several Times. I got bogged down learning frontend, designing the backend, and app requirements just to manage multi-modal input. I’m a security architect, not a web developer. I needed something that worked with the tools I already use, not another app to build and maintain.

    Here’s what I realized: so much of my work is already in the infrastructure and development space where CLI shines. And so much of my note-taking already exists in Markdown, in Obsidian. The platform is extensible. The notes are in plain text format. Claude Code can read and write files.

    The pieces were already there. I just needed to marry them.

    What I Actually Built

    Obsidian has been my note-taking home for years. Markdown files, bidirectional links, a plugin ecosystem I’ve customized over time. But it was passive - a place to store thoughts, not a system that works alongside me.

    By pointing Claude Code at my vault, I turned it into something different: an integrated system where AI agents operate directly on my knowledge base. I found a plugin by Derek Larson obsidian-claude-sidebar that neatly plugs Claude Code into my vault. Once I started refactoring all my notes my mind began to flourish with ideas. So I started building…

    Agents That Know Their Lane

    Instead of one general assistant that forgets everything between sessions, I built specialized agents:

    • Sage - The orchestrator. Manages the vault structure, tracks progress across projects, processes my inbox when I drop in voice memos and brain dumps. Sage keeps the system coherent.
    • Porty - My portfolio content partner. Drafts blog posts like this one, manages project writeups, handles changelogs and versioning. Knows my frontmatter standards and publishing workflow.
    • Coder - Development partner. Git operations, code generation, deployments. Has exclusive access to GitHub and knows where all my repos live.
    • Mentor - Career coaching. I dictate professional brain dumps and Mentor turns them into structured thinking and actionable next steps.

    Each agent has system instructions, a working directory, defined procedures. When I drop a dictation into Sage’s inbox, it knows what to do with it. When I ask Porty to draft something, it writes in my voice - or close enough that editing is minimal.

    Drop Zones for Async Work

    I over utilize dictation to such a degree that it has become a natural practice to record my thoughts however complex or disorganized and allow the AI to parse my thoughts for me. This workflow includes designated folders where I leave raw thoughts for agents to process later:

    • Career brain dumps go to /Brain Dumps/ - Mentor picks them up
    • Blog ideas land in Portfolio site/Blogs/ - Porty refines them
    • Everything else hits the root inbox - Sage routes it

    This changed my workflow more than I expected. I don’t need to be “in a session” to capture ideas. I dictate when the thought is fresh from mobile or wherever, drop it in the right folder, and the system processes it when I’m ready to engage. The capture is decoupled from the processing. Although the next step is to setup a trigger to handle that processing automatically.

    Evergreen Notes That Actually Stay Green

    My vault had years of accumulated notes. Most had no frontmatter (metadata for the non-Obsidian types). Folders were ad-hoc. Connections I’d made mentally were never made explicit in the system.

    I pointed Claude Code at the mess and let it work. In one session, it processed nearly 150 files:

    • Added consistent metadata across everything
    • Reorganized folders by semantic meaning
    • Archived what was stale
    • Preserved original content while adding structure

    I found notes I’d forgotten writing. Some of them were relevant to projects I’m working on now. That’s the evergreen promise I always wanted but never had time to maintain manually. Now it happens in the background.

    Why Not Just Live in an IDE?

    Kiro. VSCode with Copilot. Cursor. Claude Code in the terminal. These are powerful. I use them for development work. But:

    The interface matters. Obsidian is designed for thinking, not just coding. Graph views that show how ideas connect. A plugin ecosystem built for knowledge workers. Daily notes, publish workflows, templates. The community has built tools for how people actually think, not just how they code.

    Not everything is code. Career planning. Personal goals. Learning notes. Spiritual reflections. These don’t belong in a code editor. I don’t want to open VSCode to think about my five-year plan.

    Context is multimodal. My vault has PDFs, images, transcripts, reference material. The relationships between them matter. An IDE treats everything like source code.

    Claude Code operating on Obsidian gives me the power of an AI-integrated development environment with an interface built for knowledge management.

    The Bigger Picture

    This is one piece of a larger architecture I’m building. The goal: connect the systems I use today - currently manual and siloed - into something more autonomous.

    • Obsidian for knowledge management
    • GitHub for code and configurations
    • Local and remote development environments
    • Projects, Task management and calendar (eventually)
    • Discord for notifications
    • Agent handoffs and synchronous multi-system context

    All visible, trackable, verbose, dare I say… beautiful.

    The agents don’t just read and write notes. They SSH to remote servers. They manage git repos. They deploy to production. Sage coordinates; specialized agents execute. I’m building toward a system that can run multi-step workflows without me babysitting each step. Subagent and background processes for trusted repeated and low risk tasks keep my current session clean and in flow. All while stacking operations and administrative tasks that I no longer have to do.

    I spent too long trying to build this as an app. Frontend frameworks, UI components, state management - none of that is my strength, and all of it was friction. What I actually needed was a workflow that leveraged tools I already knew. Obsidian + Claude Code + clear agent definitions. That’s a start.

    Workflow Over Apps

    There’s a temptation in the AI space to build products. SaaS tools. Consumer apps. And that’s valid work.

    But as a learner and builder, I’ve found more leverage in workflow:

    • Compounding returns - Last week’s session informs today’s without me doing anything. The system remembers.
    • Personal optimization - It’s tuned to how I work, not how a product manager thinks I should work.
    • Skill development - Building this taught me things that transfer. Agent design. Prompt engineering. System architecture.
    • Immediate utility - I use this every day. It’s not a side project waiting for users. It’s my actual workflow.

    The investment compounds in ways that chasing the next app idea doesn’t.

    What’s Next

    The foundation is in place. From here:

    • More specialized agents - Fitness tracking, finance review, learning management
    • Cross-system automation - Calendar integration, task management hooks
    • Proactive context - Agents that surface relevant notes before I ask
    • Mobile capture - Seamless dictation from phone to desktop processing

    I’m still building. But I stopped copying context between sessions. The vault and the AI finally know about each other.

    That alone was worth it.

    This post was drafted in Obsidian with Porty, reviewed by Editor, and refined through iteration. The system works.