Claude Code ships with the tools it needs to work on your code: it reads files, greps them, edits them and runs git, and it does all of that well. What it does not ship with is anything that outlives the session. MCP (Model Context Protocol) is how you add that — and other things Claude Code cannot do by itself. Here's how it works, and how to tell a useful MCP server from one that only duplicates what the agent already has.
What MCP is
MCP is an open standard created by Anthropic that lets AI agents call external tools during a conversation. An MCP server is just a program that answers tool calls: it can search a database, query an API, write to a file, look up a calendar — whatever it exposes as tools. The agent reads the list of tools once, decides when to call one, and folds the result into its work.
Claude Code speaks MCP natively, and so do Cursor, VS Code, Windsurf, JetBrains, Zed and most other agents. One server, one config line per client.
The trap: tools the agent already has
The obvious first MCP server for a codebase is a search server: symbol lookup, call graphs, file reading. We built one, and learned that Claude Code already does that job well on its own: its grep and file reading are precise, and a second search tool mostly adds one more description for the model to read on every request. So we took those tools out of what the agent sees.
The rule we ended up with: an MCP server earns its place when it gives the agent something it cannot do by itself. Every tool's description is read by the model on every request, so a tool that duplicates a native one has a cost and no benefit.
What Claude Code cannot do by itself
- Remember: every session starts blank. A memory server lets the agent write what it did, why, what comes next and the traps it found — and read it back the next time.
- Talk to a person across sessions: leave a todo, ask a structured question and find the answer waiting.
- Keep a record: a timeline of every memory, note, edit, commit and lock, with who did it, that a person can read too.
- Coordinate with other agents: a lock on the file it is about to edit, so two agents on the same checkout do not write over each other; short messages between them.
- Share all of it with a team, encrypted, without a copy on someone else's server.
Reads, searches and edits your code well. Forgets everything when the session ends. Two agents on one checkout can overwrite each other. Nothing is recorded but the chat.
Still reads and searches with its own tools. Writes its memory at the end, gets it back at the start. Leaves notes and forms. Takes a lock before editing. Every event on a timeline.
How to set up Codelux AI for Claude Code
Codelux AI is a local MCP server that exposes 8 tools: memory_set, memory_get, notes_set, notes_get, code_query, code_context, code_edit, agents_msg. For Claude Code it comes as a plugin, which also brings the hooks that make memory automatic — read at session start, relevant bits added with each prompt, a reminder to save at the end — and the lock around Claude Code's own Edit and Write.
- Download the Free build from codelux.ai/download and create your account inside the app — the Free plan has no time limit.
- Run
codelux start: it serves the dashboard and the MCP endpoint onlocalhost:8766. Add your project folder from the dashboard. - Install the plugin:
claude plugin marketplace add codeluxai/codelux
claude plugin install codelux@codeluxWithout the plugin — for Cursor, VS Code or any other client — it is one line of MCP config pointing at http://localhost:8766/mcp. The agent then reads its memory when it calls memory_get or from the block Codelux AI keeps in AGENTS.md, and code_edit takes the lock for it.
Does Codelux AI work with Cursor, VS Code, and other editors?
Yes. MCP is an open standard and Codelux AI implements it — so it works with any MCP client:
- Claude Code: the plugin, with hooks
- Cursor: via
.cursor/mcp.json - VS Code: via
.vscode/mcp.json(Copilot Chat) or the Claude / Cline extension config - Windsurf, JetBrains AI, Zed: via their respective MCP config files
- Codex, Gemini CLI: the plugin ships hook configuration files for both, experimental
- Scripts and CI: via the plain REST API at
localhost:8766/api/project/{name}/...
All of these share the same memory and the same timeline, so they follow you when you switch editors.
Does my code leave my machine?
No. Memory, notes, timeline and index are stored in databases on your machine, and nothing about your code is uploaded to Codelux AI: the licence check is the only routine call. Claude Code keeps talking to Anthropic as it always does — not through us. On the Team plan, what the team shares leaves your machine only as ciphertext, encrypted with a key we never hold.
Give Claude Code a memory over MCP
Download, sign in inside the app, install the plugin — the agent writes its memory and gets it back, and everything it does is on the timeline.
Download free →← Back to the blog | Next: How to give Claude full codebase context →