Ontology Atlas MCP
MCP server keeping a Markdown map of code meaning beside the repo: the agent reads context before a task and suggests updates.
- 149 ★ on the repository
- Last commit
Extended access
Why this level
- The agent can write to the `atlas/` folder in the repository, and a connected agent sends what it reads to its own provider.
Install
- Works in
Manual install
docker run --rm -i --user "$(id -u):$(id -g)" -v /path/to/atlas:/vault ghcr.io/wlsdks/ontology-atlas-mcpThe container option from the MCP package README. The primary macOS and Windows route is the app with a bundled server from ontologyatlas.com.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent can write to the `atlas/` folder in the repository, and a connected agent sends what it reads to its own provider.
This is third-party code. Review the repository files before installing.
What it does
Ontology Atlas keeps a map of the system in the repository's `atlas/` folder: one Markdown file per concept, with code paths, dependencies and open questions. The `ontology-atlas-mcp` server reads and writes that folder directly on disk, even with the app closed. Before a task the agent gets a compact `agent_brief`, and after it suggests updates. Proposals stay proposals until a person accepts them as a Markdown diff. The map is shown by a macOS app and a browser version.
Who it is for. Teams where agents change code fast and nobody holds the whole picture: team leads and developers of large repositories.
Good fit when
- Before a complex task the agent needs a map: what this part of the code is, which files and dependencies.
- After a change the agent suggests a description update and the developer reviews the diff in Git.
- You need to see what a change may have touched and separate known from unknown.
- A new developer reads the same map as their agent.
Not a fit when
- The repository is small and its structure is obvious.
- You need runtime-impact guarantees; a line on the map is a declared relationship, not proof.
- There is no Node or container and the app cannot be installed.
Example request
Read the Atlas map for the payments task, show the affected concepts and open questions, then propose a description update without writing it.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
OATLAS_VAULTrequired- Absolute path to the vault folder holding the Markdown map
Limitations
The project is early: by the author's own benchmark, answers with Atlas are not yet better and work got slower. There are four install channels and npm is not one of them: the app, a source checkout on Node 24 and pnpm, an `.mcpb` file and a container. The Windows app is an unsigned beta, and Linux uses the browser or source. A connected agent sends what it reads to its own provider, which the Atlas log does not cover.
How to disable. Remove the `ontology-atlas` entry from `.mcp.json`, `.codex/config.toml` or `.cursor/mcp.json`. The `atlas/` folder stays plain Markdown in the repository; delete it or keep it as documentation.
FAQ
How do I connect an agent?
In the app open Agents, MCP and press Connect for your tool. The button writes a config file: `.mcp.json` for Claude Code, `.codex/config.toml` for Codex, `.cursor/mcp.json` for Cursor.
Can I skip the app?
Yes. The agent starts the MCP server itself and it works on the folder on disk without the app. Container run: `docker run --rm -i --user "$(id -u):$(id -g)" -v /path/to/atlas:/vault ghcr.io/wlsdks/ontology-atlas-mcp`.
Does a human approve every agent write?
Not necessarily. The docs caution that an MCP connection does not guarantee human approval of every write; behavior depends on client settings.
How many tools does the server have?
The package description lists 40: 24 read and 16 write.
Where does the map live?
In the repository's `atlas/` folder: one Markdown file per concept. It is cloned, branched and reviewed with the code.
Discussion
No questions or reviews yet.
Related
MCP serversA local repository map for agents: the atlas CLI and MCP find the relevant code by file purpose instead of opening everything.
One MCP server gives agents a code map across over 300 languages, git history, shared memory and document search with OCR.
An MCP server and CLI in Go: the agent keeps an index of code and database schema in the repository, versions it in Git and reads it whole before each task
Tasks, docs, specs and decisions live as markdown files in your repository, and the agent reads them through the CLI and MCP
An MCP server that graphs repository dependencies, writes docs from the graph and answers agent questions via codebase-qna
Local code structure graph: builds a repository map with tree-sitter and serves the assistant only the context a change actually touches over MCP
README in short
The code map lives in Markdown beside the code and is reviewed in Git. The agent reads it over MCP, and a person sees the same map in the app or the browser.