Why Obsidian became the central brain of every engagement

On every Twill engagement we ship an Obsidian vault. Decision logs, runbooks, codebase notes, conventions — all of it lives in a folder of Markdown files we hand off on Day 21. Why a 2026 dev agency standardizes on what is nominally a "note-taking app" deserves an explanation. So here it is.

The numbers

Start with the data. As of February 2026, Obsidian sits at 1.5M+ monthly active users, growing 22% year-over-year, with roughly $25M ARR and zero VC funding. That's up from the ~1M baseline reported in 2023. The community ships more than 2,000 plugins through the official directory. The Discord has crossed 110,000 members.

Market share is the part most people misread. Obsidian holds roughly 8% of the broader PKM category — small in absolute terms, dominant in the segment that matters: developers, researchers, technical writers, anyone who works in plain text by default.

Notion still leads in search volume, but the steepest growth curve in the category belongs to a tool that ships zero servers and stores your data as files on your disk.

The market context

Knowledge management software is on track for $16.2B in 2026, projected to reach $74B by 2034 at a 13.8% CAGR — driven mostly by generative-AI integration. The personal knowledge management subset alone is expected to hit $4.2B by 2028.

Most of that growth is captured by tools that aren't local-first. Notion dominates team collaboration. Roam Research has been in slow decline since 2020. Logseq carved out a small open-source niche. Apple Notes wins the casual market by being preinstalled.

Obsidian is winning the segment that matters most to a working dev team — the segment that wants files it owns, edited by tools it chooses, readable by every program it already runs.

How the field stacks up

The PKM landscape, ranked on the dimensions that matter to AI work.

Tool	Storage	Lock-in	Plugin ecosystem	Team collab	AI-friendly (file access)	Best for
Obsidian	Local Markdown	None	2,000+	Weak	Yes	Individuals, devs, researchers
Notion	Cloud DB	Heavy	Built-ins only	Strong	No	Teams
Logseq	Local Markdown	None	Smaller	Weak	Yes	Open-source devs
Roam Research	Cloud	Medium	Smaller	Weak	No	Power users (declining)
Apple Notes	Proprietary cloud	Heavy	None	Weak	No	Casual

Why local-first won the AI era

The structural argument is simple. AI agents need file access, not API access.

A folder of Markdown files is the lowest-common-denominator interface in software. Claude Code reads it. Cursor reads it. Every editor, every grep, every script you've ever written reads it. There is no integration layer to build, no rate limit to negotiate, no auth flow to maintain. The agent opens a file, the agent edits a file, the agent moves on.

Compare that to a cloud-database PKM. To let an agent see your notes you'd need to build or maintain an MCP server, manage credentials, deal with pagination, hope the schema doesn't change under you, and accept that latency now lives between your model and your knowledge. None of that is fatal — but every layer is a thing that can break, and a thing somebody has to own.

Native MCP integration shipped in late 2025 and was production-grade by early 2026, which made "vault as agent context" a first-class pattern. The vault becomes part of the agent's working memory. It knows your project's conventions, your past decisions, your codebase notes, the failure modes you hit last sprint. The same files your humans read.

No lock-in also means model-agnostic. Claude, GPT, Gemini, local Ollama — they all work the same way against the same vault. The vault outlives the model.

How we use it

The engagement-level workflow is deliberate and boring, which is the point.

Day 1. Spin up the project vault. Folders for decisions, runbooks, codebase notes, failure modes, weekly reviews. We seed it with a starter structure tuned to the engagement.

Across three weeks. Every architectural decision, every failure mode, every "you'll need to know this on Tuesday morning at 6:14am" — it goes in the vault. We pair the vault with Claude Code so the agent reads from it as context for every change. The vault and the codebase grow together; neither leads the other.

Day 21. The vault gets handed off. The team owns it. Future engineers onboard from it. Future agents read from it. The workflow keeps working because the vault keeps existing — independent of any agency, any model, any subscription.

What the numbers don't show

Obsidian isn't perfect, and pretending otherwise would be the kind of marketing puff we try to avoid.

Team collaboration is genuinely weaker than Notion. Multi-user editing is a Sync add-on, not a first-class feature. We work around this by treating the vault as single-source-of-truth for durable knowledge and using the team's existing tools — Linear, Slack, GitHub — for coordination and ephemeral chatter. The vault is for things that should still be true in six months.

The plugin ecosystem is also a maintenance surface. 2,000 plugins is a lot of community code, and not all of it is healthy. We pick a small, vetted set per engagement and freeze it. Updates are intentional, not automatic.

The learning curve for power features — Dataview, Templater, links, the graph view — is real. We don't ship the full studio setup. We ship a usable starting point a non-expert can extend.

We pick Obsidian because the structural fit with AI agents outweighs the rough edges. That math has only gotten clearer over the last two years.

Closing

The note-taking app didn't win because of features. It won because plain Markdown plus local files happened to be the right protocol for an AI-augmented workflow. The next time you read a hot take about which PKM is the future, check whether the author's tool can be read by a grep command. If not, it isn't.

If you want to see how Obsidian fits into the rest of our toolchain, the pipeline lives here.