Keep understanding your system as agents change its code. Before accepting a change, you need to know which capability it serves, what rules matter, and what still needs inspection. Source code, documents, and people can supply different parts of that evidence; the producing agent's summary is a claim to review.
Ontology Atlas captures that layer of meaning in a single folder of Markdown files. The file frontmatter serves as nodes and relationships. No separate database, login, or server required.
A useful analogy is an IDE for codebase meaning. Atlas brings construction, inspection, validation, review, and maintenance of the codebase ontology into one workbench; it does not edit, build, run, or debug source code.
Two types of readers
- Humans: Inspect recorded meaning, its evidence, and unknowns through the map, documents, and change review. Correct or reject a proposed meaning; acceptance is separate from code review, merge, or deployment.
- AI Agents: Read and write the same files via MCP. Task-aware context can provide starting points, domain context, implementation evidence, declared impact boundaries, and verification paths for source inspection.
Atlas is one meaning layer maintained by both. You need not open the app for every task; an agent can consume the context over MCP, and you can open the workbench when you need to inspect or correct it. A connection does not guarantee that every host uses the context or applies the same review permissions.
When to use it
- When the codebase grows so large that "why is this like this?" becomes an archaeological dig every time
- When you're pasting the same background context into AI agent sessions repeatedly
- When planners and developers use the same words to mean different things
- When you need to inspect declared dependencies, evidence, and unknown impact before fixing something
When not to use it
- Projects with just a few files: everything fits in your head
- Cases needing only ordinary note storage: the ontology workflow may add no value; Atlas Library can also hold source documents without ontology nodes
- Questions where you only need to know code structure: grep and language servers are faster
Atlas complements source search, CodeGraph, and language tools. It preserves reviewable meaning linked to implementation evidence. Reliable reconstruction across unfamiliar legacy repositories and a complete task-bound Meaning Diff remain development work; a graph cannot guarantee a safe code change.