Ontology Atlas keeps a map of what a codebase does and why as Markdown files inside the repository, where both people and coding agents read and update it. Each concept (a project, domain, capability, or element) is one Markdown file whose frontmatter names the code it describes and its dependencies on other concepts, so the map is versioned and reviewed alongside the code.
A coding agent reads the concepts, code paths, dependencies, and open questions a task touches before it starts, and proposes updates afterward. Proposals stay proposals until you accept, correct, or reject each one as a Markdown diff. The map treats a line between concepts as a declared relationship rather than proof of runtime impact, and shows missing evidence as unknown.
Agents reach the folder through an MCP server that reads and writes it on disk even when the app is closed, or through the Agent Client Protocol, which runs Claude Agent and Codex in the app's own chat with each write waiting for approval. Claude Code, Codex, Cursor, and Antigravity connect over MCP.
Atlas ships as a macOS app, a browser version, and a CLI. It has no backend, account, or telemetry, and model calls with your own API key or a local model are opt-in and logged with their destination. The project describes itself as early: its own benchmark has not yet shown better agent answers with Atlas, and runs with Atlas were slower.
Features
- Markdown ontology: one file per concept, with a stable ID, current address, and the code path it describes
- Topology map: projects, domains, capabilities, and elements with their dependencies, plus each project's code-evidence state
- Agent context over MCP: Claude Code, Codex, Cursor, and Antigravity read the concepts and open questions a task touches
- Reviewed agent proposals: agent edits stay proposals until accepted, shown as the exact Markdown diff before saving
- In-app agents: Claude Agent and Codex run in the app's chat over the Agent Client Protocol, with each write approved
- Library: keeps source documents as they arrived and writes wiki pages from them with every fact cited
- Insights: lists what to fix next, grouped by measurement rather than a score
- Analysis: compares recorded dependencies between responsibilities and opens the declaring document beside a claim
- Harness view: shows what the repository tells, blocks, and watches for its coding agents
- Automations: scheduled checks that keep the concept folder current
- Local-first: no backend, account, or telemetry, with opt-in model calls logged by destination
