// blog / guide

Prompt management for Claude Code with MCP: setup, slash commands and @-resources

Pasting a prompt into Claude Code by hand is the slow part. With an MCP server, the agent can find, fill and run your saved prompts itself, and save good ones back.

Set it up

Install the CLI and save a prompt (no account yet).

terminal

npm i -g promptlike
ple save "code review" -m "Review this diff for bugs: {{diff}}" -t review

Sign in. The MCP server reads your account library, so this step is required.
terminal
```
ple login
```
Upload what you saved locally. Signing in never uploads anything by itself.
terminal
```
ple sync --dry-run   # preview
ple sync             # one-way upload; local files are kept
```
Add the server to Claude Code.
terminal
```
claude mcp add promptlike -- npx -y promptlike-mcp
```
Use it. In a Claude Code session, run a saved prompt as a slash command, or attach one with @.
claude code
```
/mcp__promptlike__code-review
```

The package is promptlike-mcp on npm. The full reference, including the Cursor and Codex setup, is in the MCP docs.

What Claude Code can do with it

mcp tools

list_prompts, search_prompts   find a prompt, with its {{variable}} names
get_prompt                     fetch one, optionally a pinned version
fill_variables                 fill the {{variables}} and return ready text
suggest_prompts                up to 3 of your prompts that fit a task
save_prompt, update_prompt     add a prompt, or add a new version

Each saved prompt is also an MCP prompt, so its {{variables}} become slash command arguments, and a resource you can attach with @. A built-in /mcp__promptlike__save-this asks the agent to pull the reusable instruction out of the conversation, show it to you and save it.

What it will not do

Delete. There is no delete tool; an update is a new version and the old one stays.
Guess. An update needs the exact title or id, and a near match gets a "did you mean", never a write.
Store secrets. Content that looks like an API key or token is refused; use a variable instead.

Because an agent can add versions, the history is what keeps that safe: you can diff what it saved against the version before, and roll back. For the general case, see prompt version control.

Where the library fits next to CLAUDE.md

Rules that should apply to every session belong in CLAUDE.md, kept short (shrink your CLAUDE.md). Task prompts you reuse, with variables and a history, are what the library is for. How the options compare is in best prompt manager for Claude Code.

FAQ

Do I need an account to use the MCP server?

Yes. The server reads your account library using the session that ple login stores. The CLI on its own works without an account, but the MCP server does not.

Can Claude Code change or delete my prompts?

It can add: save_prompt creates a prompt and update_prompt saves a new version, with the old one kept. There is no delete tool, and content that looks like an API key or token is refused.

Which MCP clients does this work with?

It was tested in Claude Code. The server speaks standard MCP over stdio, and the docs include setup steps for Cursor and Codex written from their own documentation, which we have not yet run in a live session.

Does my prompt text go to a third party when Claude Code reads it?

The server fetches your prompts from your PromptLikeEngineer account and hands them to the agent you are already running. Which model processes the text is decided by that agent, not by the server.

How is this different from CLAUDE.md or slash commands in .claude/commands?

Those are files in one project or your home folder. A prompt library over MCP adds one history and search across projects, with variables filled for you. Keep always-on rules in CLAUDE.md and reusable task prompts in the library.

Try it before you sign up

See the version history the agent writes into, before you connect anything. It runs in your browser with no account, and nothing is sent to our servers.

Try it in your browser Or create a free account