Project .mcp.json for Claude Code with filesystem, memory, fetch and sequential-thinking servers
A project-scoped .mcp.json that connects Claude Code to four official reference MCP servers: filesystem access to chosen folders, a persistent knowledge-graph memory, web page fetching, and a structured step-by-step thinking tool.
File in Engineering by AISkillHub Library
What it does
A `.mcp.json` file at the root of a repository gives everyone who opens the project in Claude Code the same set of MCP servers, without each person running `claude mcp add` by hand. This file sets up four servers from the official Model Context Protocol reference implementations, all of which run locally and need no API keys.
**Servers included**
| Name | Package | What it adds | |---|---|---| | `filesystem` | `@modelcontextprotocol/server-filesystem` (npm) | Read, write, search and list files, limited to the directories passed as arguments (`./docs` and `./data` here) | | `memory` | `@modelcontextprotocol/server-memory` (npm) | A local knowledge graph of entities, relations and observations that persists across sessions | | `sequential-thinking` | `@modelcontextprotocol/server-sequential-thinking` (npm) | A tool for breaking a problem into explicit, revisable steps | | `fetch` | `mcp-server-fetch` (Python, run with `uvx`) | Fetches a URL and converts the page to Markdown |
**When this is useful**
- The filesystem server is most useful for folders **outside** the project, such as a shared specs directory, because Claude Code already works with files in the project itself. Point it at exactly the folders you want exposed. - Memory helps on long-running projects where you want facts (decisions, naming conventions, contacts) to persist across sessions. - Fetch lets Claude read documentation pages and convert them to Markdown.
**Security notes**
- Claude Code asks each user to approve project-scoped servers from `.mcp.json` before using them. Review this file in PRs like code, since it defines commands that run on developers' machines. - The fetch server can retrieve any URL the machine can reach, including internal addresses. Remove it on machines with access to sensitive internal services, or restrict network access. - Fetched web pages are untrusted input and can contain prompt injection.
**What it will not do**
It does not include servers that need credentials (GitHub, Slack, databases). Add those yourself with environment variables, never with tokens written into this file, since it is committed to the repo.
How to use it
1. Requirements: Node.js with `npx` for the three npm servers, and `uv` (which provides `uvx`) for the fetch server. Remove the `fetch` entry if you do not want to install `uv`. 2. Copy `.mcp.json` to the root of your repository. 3. Edit the filesystem arguments: replace `./docs` and `./data` with directories that exist and that you want to expose. Relative paths are resolved from where Claude Code is started (normally the repo root). Absolute paths work too. 4. Remove any servers you do not need. 5. Start Claude Code in the repository. It asks you to approve the project servers; approve the ones you want. 6. Check status with the `/mcp` command inside Claude Code, or `claude mcp list` from the terminal. 7. Commit the file so teammates get the same setup. Never put tokens or keys in it.
Files
Save each file at the path shown, relative to your project root.