promptlike-mcp

promptlike — Claude Code plugin

Reach your PromptLikeEngineer prompt library from inside Claude Code. Your prompts stay versioned where you saved them; Claude pulls one by name at the moment it needs it — no copy-paste, no stale duplicates in a system prompt.

Install

From this repo's own marketplace — works today:

claude plugin marketplace add zeynepaslierhann/promptlike-plugin
claude plugin install promptlike@promptlikeengineer

Inside a Claude Code session the same two steps are /plugin marketplace add zeynepaslierhann/promptlike-plugin and /plugin install promptlike@promptlikeengineer.

From the community marketplace (once this plugin is listed there):

/plugin marketplace add anthropics/claude-plugins-community

Then install promptlike from @claude-community.

Or straight from a clone of this repo, handy for trying it without adding a marketplace:

git clone https://github.com/zeynepaslierhann/promptlike-plugin
claude --plugin-dir ./promptlike-plugin

Sign in once so the server can read your library (it shares the CLI session):

npm i -g promptlike && ple login

Without signing in the plugin still loads and its tools are listed — each call just answers "Not logged in. Run ple login in a terminal first (install: npm i -g promptlike)." instead of returning prompts. The suggestion hook then uses your local prompts (ple save).

What it adds

Component	What it is
promptlike MCP server	Tools: list_prompts, search_prompts, get_prompt, fill_variables, suggest_prompts (read) · save_prompt, update_prompt (write — add only)
Your prompts as slash commands	Every saved prompt is an MCP prompt: type / and pick it, or run /mcp__plugin_promptlike_promptlike__<prompt-name> (a title such as "Code Review" becomes code-review). Its {{variables}} become arguments. /mcp__plugin_promptlike_promptlike__save-this saves the reusable prompt from the current conversation
Your prompts as @ resources	promptlike://<repo>/<prompt> — attach a saved prompt with @
Suggestion hook	Before each message, a local match against your saved prompts; when one fits, Claude gets a short note (see below)
/promptlike:prompt-library skill	Teaches Claude to reach for a saved prompt instead of writing one from scratch, and to save good ones back

The plugin starts the server with npx -y [email protected], pinned to that exact version.

Examples

Each example starts with something you do. Swap in the names of prompts from your own library.

1. Ask for a prompt by name. You type:

Run my code-review prompt on the changes in this branch.

Claude looks the prompt up (get_prompt, after a search_prompts if the name is not exact), fills its {{variables}} with fill_variables, and follows the result. If the name matches several prompts, Claude shows you the candidates instead of guessing.

2. Run a prompt from the / menu. You type:

/mcp__plugin_promptlike_promptlike__code-review src/auth.ts security

The words after the command fill the prompt's {{variables}} in the order they first appear in it (for a prompt that mentions {{file}} and then {{focus}}, the two words above). A variable you leave out stays as {{name}} in the text Claude receives.

3. Keep something you just worked out. You type:

/mcp__plugin_promptlike_promptlike__save-this

The prompt tells Claude to pull the reusable instruction out of the conversation, mark what changes between uses as {{variables}}, show you the title, description and text, and then call save_prompt. Saving is add-only, and text that looks like an API key or token is refused.

Saving from Claude Code

Ask Claude to save a prompt, or run /mcp__plugin_promptlike_promptlike__save-this. Writes are add-only: save_prompt creates a prompt, update_prompt adds a new version and keeps the old one. Nothing is ever deleted, and saving the same thing twice changes nothing. Plan limits apply exactly as on the web — when you hit one, Claude tells you and nothing is saved. Content that looks like an API key or token is refused.

Suggestion hook — what it does and how to turn it off

On every message you send, Claude Code runs scripts/suggest.sh, which runs ple suggest --hook (a UserPromptSubmit hook). It:

• matches your message against your saved prompts on this machine — a local copy of your library (or ~/.promptlike when signed out). Your message is not sent anywhere;
• stays silent for slash commands, short messages (under 20 characters), weak matches, and prompts it already suggested in this session;
• otherwise adds one short note for Claude naming up to three prompts. Claude fetches one only if it is clearly relevant;
• needs the CLI (npm i -g promptlike); without it the hook does nothing. It never blocks your message: it answers in well under a second or stays silent.

Turn it off any of these ways:

• /config → PromptLikeEngineer → Suggest saved prompts → false (or run /plugin configure promptlike@promptlikeengineer)
• ple config set suggest off
• PLE_SUGGEST=off in your environment (or in Claude Code's settings.json → env)

How often a suggestion is shown is counted on your machine only (ple stats shows it); the server only learns when Claude actually fetches a suggested prompt.

Data and privacy

This is everything the plugin runs and every place it talks to.

Suggestion hook. Before each message Claude Code runs scripts/suggest.sh, which runs ple suggest --hook if the ple CLI is installed (otherwise nothing happens). Matching runs on your machine. The hook does not send your message anywhere, makes no network request itself and never waits for one. One thing does happen in the background: if you are signed in and the local copy of your library is missing or more than 10 minutes old, the hook starts a detached ple suggest --refresh-cache (at most once a minute) that downloads your own prompt list from promptlikeengineer.com with your ple login session, so the next match uses a fresh copy. Each of the three off switches above stops the hook, and that refresh with it. When a prompt fits, Claude gets a two-line note naming up to three of your prompt titles; like any text in your conversation, it is sent to the model with that turn.

Files on your machine. The CLI keeps these in ~/.config/promptlike/ (or the folder named in PROMPTLIKE_CONFIG_DIR), created with owner-only permissions where the system supports them: config.json (your session: tokens, email and user id, written by ple login), settings.json (the on/off switch), cache/library.json (up to 500 prompts: title, description, tags, repo, use count and the first 600 characters of the text) and cache/suggest-state.json (a counter and the prompts suggested in your last 20 Claude Code sessions). Signed out, the hook reads ~/.promptlike instead.

MCP server. Claude Code starts it with npx -y [email protected]; the first time, npm downloads that exact version from the npm registry. It runs on your machine and talks only to https://promptlikeengineer.com (or the address in PROMPTLIKE_API_URL, if you set one), authenticating with the session ple login stored and refreshing it through the same host when it expires. It has no other network code. What it sends:

• reads (list_prompts, search_prompts, get_prompt, fill_variables) ask your account for your prompts; Claude Code also asks for the list that fills the / menu and the @ suggestions (up to 500 prompts, kept for 30 seconds);
• when Claude fetches a prompt — by tool, from the / menu or as an @ resource — the server records a use event: which prompt and version, which of those three ways it was reached, whether it was offered as a suggestion, and the MCP client's name and version (for example claude-code/2.1.179);
• suggest_prompts sends the task description Claude gives it (up to 2,000 characters) so the server can match it against your library; the endpoint does not log or store that text. The hook above never does this;
• save_prompt and update_prompt send the text you asked to save.

Limits on what it does. Writes are add-only: the server has no delete tool, saving the same text twice changes nothing, and content that looks like a secret (API keys, tokens, private keys) is refused — the error names the kind, never the value. Prompts you save live in your PromptLikeEngineer account; how the service handles them is in its Privacy Policy and Terms.

Troubleshooting

What you see	What to do
"Not logged in. Run ple login in a terminal first (install: npm i -g promptlike)."	Run npm i -g promptlike && ple login once, then ask again. Tools use the new session on the next call; slash commands and @ resources are listed when the server connects, so start a new Claude Code session if they are missing
"Session expired. Run ple login in a terminal to sign in again."	Run ple login again
"Rate limited. Retry in 12s."	Wait that long, then ask again
"Cannot reach https://promptlikeengineer.com. Check your connection." or "No response from https://promptlikeengineer.com within 15s."	Check your connection and try again
"… Plan limit reached — nothing was saved. Upgrade: …"	Your plan's limit stopped the save and nothing changed
claude mcp list shows plugin:promptlike:promptlike as failed to connect	The server needs Node.js 18 or newer, npx on your PATH and, the first time, access to the npm registry. Run npx -y [email protected] in a terminal: it prints promptlike-mcp v0.5.0 ready (stdio). when it starts (stop it with Ctrl-C), or the error if it cannot
No suggestion notes appear	The hook needs the ple CLI on your PATH. It stays silent for slash commands, messages under 20 characters, weak matches and prompts already suggested in this session, and when it is switched off: ple config get suggest shows whether it is on

Self-hosting

Point at another deployment with PROMPTLIKE_API_URL (the same environment variable the CLI honors).

Support and security

Questions and bug reports: open an issue at https://github.com/zeynepaslierhann/promptlike-plugin/issues or write to Zeynep Aslı Erhan · [email protected]. Please report a security problem by email rather than in a public issue. This repository holds only the plugin files; the MCP server is the npm package promptlike-mcp and the CLI you sign in with is promptlike, both MIT licensed.

Links

• Product: https://promptlikeengineer.com/?utm_source=github&utm_medium=readme&utm_campaign=ple_dist
• Documentation: https://promptlikeengineer.com/docs/mcp
• Privacy Policy: https://promptlikeengineer.com/privacy · Terms: https://promptlikeengineer.com/terms
• MCP server on npm: https://www.npmjs.com/package/promptlike-mcp
• CLI on npm: https://www.npmjs.com/package/promptlike

MIT licensed.