Skip to main content
Your knowledge is markdown in a git repo you own. Isomorphic is the layer that lets an LLM maintain it: a Model Context Protocol server with a librarian’s toolkit, a content index that keeps reads fast at any size, computed views, and an in-client viewer and WYSIWYG editor so a non-technical teammate never has to open GitHub.
The Isomorphic page viewer

The page viewer: frontmatter as properties, a table and count computed from the pages that link here, and the linked references below.

Getting started

Connect Claude to the hosted service and create your first brain.

Self-hosting

Run it yourself, from a folder on your laptop to a multi-tenant deployment.

Architecture

How a tool call travels through the system, file by file.

Building with Claude Code

The setup that lets Claude Code write most of this codebase.

​
Try it in two minutes, with no accounts

The real MCP server, the real librarian tools, and the real content index, against a git repository on your disk. No GitHub account, no Cloudflare account, no tokens. Connect it with claude mcp add --transport http isomorphic-local http://127.0.0.1:8788/mcp, or open http://127.0.0.1:8788/b/local/notes in a browser for the same viewer and editor with no MCP host at all.

​
Why not a folder of markdown and a coding agent?

You can point Claude Code at a folder of notes today, and for one person that is a fine place to start. Isomorphic is for what goes wrong after that:
  • Moving a page breaks every link to it. move_page repoints each inbound link, markdown and wikilink alike, in the same commit. delete_page tells you what still points at the page.
  • An agent rewrites a whole page to change one paragraph. write_page takes exact find-and-replace edits and appends. An anchor that matches zero or several times aborts the whole call, so a batch is never half-applied and a page is never silently clobbered.
  • Reads stop scaling. A derived content index keeps search, backlinks, validation, and the graph to one or two queries on a 3,000-page brain, and checks the branch HEAD on every read so an edit made on github.com or by another agent is never served stale.
  • Listings drift. A fenced okf-view block is a listing, table, or count computed from backlinks or frontmatter, recomputed on every read instead of maintained by hand.
  • Half the team will never open a terminal or GitHub. The viewer and editor render inside the conversation and in a browser tab. Teammates sign in with an email link, nobody needs a GitHub account, and orgs, roles, invitations, and per-brain sharing are all built in.

​
What works where

The server never calls a model. It is an MCP server; the client brings the model, and any client that speaks the protocol works.

​
Why this exists

Writing knowledge down is a separate job from doing the work, which is why it does not get done. LLMs are good at that separate job and bad at doing it into a database whose shape they cannot see. So the substrate is one markdown file per concept, in a normal git repository, in the Open Knowledge Format. You can read it on github.com, edit it in Obsidian, grep it, diff it, and review a change in a pull request. If you stop using Isomorphic tomorrow you still have everything.
Open source under AGPL-3.0-only. Run it, fork it, deploy it for your own company, sell services around it. If you modify it and let others use your version over a network, you owe those users your changes. Your knowledge base is your data and the license does not reach it.