Skip to main content

v1.7.0codebase ontology · open source

Give Claude Code or Codex task context over MCP from one shared Markdown vault. Inspect recorded capabilities and rules, their evidence, and what remains unknown. Decide which proposed meaning changes to keep.

Download for Apple SiliconWindows x64 betaunsignedPlayground

Signed and notarized by Apple · no Atlas backend

Version
v1.7.0 · Oct 10, 2026
Requires
macOS 12 or later · Apple Silicon
DMG
88.3 MB
SHA-256
df4cfdb2…df27aadc
ChangelogWhat changed in v1.7.0
Sourcewlsdks/ontology-atlas

Illustration · concept names and files from this repository's ontology

An illustration. Files in this repository rise into its ontology: Ontology Atlas, its domains, and capabilities and elements that each point at a file. A coding agent asks Atlas over MCP with get_concept for capabilities/mcp-tool-server. Atlas answers with MCP tool server (Agent access) and the declared relations the answer touched. It depends on Construction guidance, Vault validation, Import dependency inference, and Task agent brief; In-app coding agent session depends on it. The agent then proposes one connection: MCP tool server → Meaning write safety. A person allows it once before it is written. The names, files and relations come from this repository's own ontology; the exchange itself is illustrative.

Demo

Inspect the context your agent reads

This recorded demo selects a concept and its evidence, then Codex follows a declared path in the same vault. It shows a bounded lookup, not a complete change-impact check.

What Codex was asked in this take
Use only get_concept and find_path to explain the path from capabilities/mcp-tool-server to capabilities/ontology-map in one sentence. Do not change files.

28s from one take. Only the agent's thinking time is cut; nothing is sped up, and there is no sound.

After a commit

When your agent commits, you see which meanings to recheck

Concepts on the map name the files they describe. When a commit in your Git repository changes one of those files after its concept was last written, the app's Insights lists that concept with the file and both dates. Reread it yourself or ask your agent to; its answer waits for you.

Illustration · concepts and files from this repository's ontology; the commit is an example

An illustration. A coding agent's commit changes three files in this repository: src/views/library/ui/LibraryPage.tsx, src/features/library/lib/wiki/judge-page-write.ts, and src/features/saved-constellations/model/use-saved-constellations.ts. Each is the file a concept on the map names: Library workspace, Wiki page write judge, and Saved constellations. A light runs from each changed file to its concept, the concept takes the amber mark Insights gives a meaning whose code moved, and each file's path moves into the Insights brief, which reads “3 concepts whose meaning stands while the code under it moved” and offers “Ask the agent”. The concepts and files are this repository's own; the commit is an example.

Screens

One folder, read seven ways

The same Markdown folder opens as seven screens in the app. Press a name to see that screen; Map opens live in this browser.

Map
The Harness screen of the Atlas app

This repository's own agent setup, read from its files: what agents are told, what stops them, what runs afterwards. Captured from the app.

Open Harness

After you install

Open your project, and let your agent draft the map

The app opens by asking for a folder. Open your project's, and the first card offers to have your coding agent read the code and draft the map. Nothing is written until you allow it, and what is written is Markdown in that folder, never your code.

Illustration · the app's own screens and words, drawn small

An illustration in three steps. First, the app's first screen, “Open a folder to start working”, with the button “Open my folder”. Second, a code folder is open and the first card reads “Let AI draft the map from your code”: It reads this folder and proposes the projects, domains and capabilities it finds. Nothing is written until you approve. What it writes is Markdown in this folder, never your code. Its button “Ask the agent” hands the request to the coding agent. Third, the draft arrives as one card, “Creates 12 concepts”: Allowing writes it into the Markdown files in this folder. Nothing changes before then. After “Allow once”, the card reads “Allowed once. The agent carries on.”, 12 concepts fill the map, and Insights lists them as “12 concepts an agent wrote that no person has looked at”.
GuideChangelog
MIT licensedThird-party licensesLocal-first · Next.js · TypeScript · Canvas 2D · MCP

Claude Code, Codex, Cursor and the other product names here are trademarks of their respective owners. Ontology Atlas is not affiliated with or endorsed by any of them; the names say what it connects to.