Skip to content

New PromptVibe Connector

Write better prompts inside the AI agent you already use.

Connect PromptVibe to Claude, Claude Code, Codex, Cursor, Gemini, Grok or ChatGPT. Name the tool and get a prompt with scope locks, stop conditions and a clear definition of done. A vague request gets two or three quick questions first, and the finished prompt goes straight to your agent to run — over a connection you can cut off in one click.

  • Keys stored only as hashes
  • OAuth 2.1 sign-in with PKCE
  • Revoke any time, instantly
  • Secrets stripped from prompts
Use PromptVibe to diagnose this prompt for Claude Code: “fix the login bug”
promptvibe › diagnose_prompt
18/ 100 — your agent would have started editing files on this.
  • highVague task or a very short request
  • mediumNo definition of done
  • highNo output format or length
  • highNo file, function or scope boundary
Real output of the diagnose_prompt tool. It runs instantly and calls no AI model.

Why connect it

Agents act on what you tell them. Tell them better.

A chat model answers a vague prompt with a vague answer. A coding agent acts on it — it edits files and runs commands — so the prompt matters more, and it is the part most people write in a hurry. The connector puts a careful prompt writer one sentence away, inside the tool you are already in.

Stays in the window you are working in

Ask for the prompt where it will run. If your request is vague it asks two or three quick questions, once. Then your agent runs the finished prompt — no second tab, no copy, no paste.

Shaped for the tool you name

25 tool routes cover chat models, coding agents, image, video, voice and automation tools. A Claude Code prompt and a Midjourney prompt need different things — you get the right one.

Catches the mistakes that burn runs

diagnose_prompt checks for a vague task, no scope, no stop condition and more — instantly, with no AI model and no credits.

Learns your standing rules

Correct the same thing in two separate prompts and it becomes a rule — “never touch the migrations folder.” Answer a follow-up question the same way twice and it stops asking. Plain counting, no hidden model, and you can read and erase it.

One setup, every agent

Your preferences live in your PromptVibe account, not in one app’s config, so they follow you from Claude Code to Cursor to Codex.

Nothing to install

It is a remote server. For most apps you add a URL and sign in; there is no package and no process running on your machine.

What the instant check looks for:

  • Contains what looks like a credential
  • Vague task or a very short request
  • No definition of done
  • No output format or length
  • Describes the feeling instead of the fault
  • Asks for the whole thing at once
  • Relies on context the model does not have
  • Invites made-up facts or citations
  • Asks for private reasoning
  • No file, function or scope boundary
  • Agentic task without stop conditions
  • Image prompt with nothing excluded

Setup

Connect your agent in about a minute.

Pick your app. Apps that can sign in need only the server URL — you approve the connection in a PromptVibe window. The rest get a connector-only key, created for you in one click and pasted into the command.

Claude Code

Sign inAPI keyMatches vendor docs

One command, then sign in.

Steps

  1. Run the command below in your terminal.
  2. Start Claude Code, type /mcp, choose promptvibe and sign in with your PromptVibe account.
claude mcp add --transport http --scope user promptvibe https://www.promptvibe.co/api/mcp

Check it works. Ask: “Use PromptVibe to write a prompt for Cursor that adds rate limiting to this API.”

Prefer a key? Add --header "Authorization: Bearer YOUR_KEY" to the command; the Codex and Gemini CLI setups above can create one for you in a click.

Making a key by hand instead? Use your account page; each key is shown once, so paste it straight into the config.

Sign in (OAuth)

For apps that can sign in: Claude on the web, desktop and mobile, ChatGPT, Claude Code, Cursor and VS Code. You give the app the server URL, it opens a PromptVibe window, you press Allow. The access it receives is a connector-only key you can see and revoke on your account page.

API key

For coding assistants and the APIs: Codex, Gemini CLI, Windsurf, Cline, Grok and OpenAI. The client sends the key as Authorization: Bearer …. Only a hash is stored, you can have several, and revoking one cuts it off at once.

Security & privacy

How secure is it?

Short answer: your agent gets a connector-only key that you can revoke in one click — not your password, not your billing, not your files. Here is what protects you, what we store, and the one trade-off to know about.

Keys are never stored in readable form

Each key (pv_live_…) is 256 bits of randomness. We keep only its SHA-256 hash and show the key once, so a database leak cannot yield a working key.

Connector-only access

Keys made by signing in, or by the one-click buttons, work on the connector and nowhere else. Every other PromptVibe API refuses them, so they cannot change your plan or billing.

Sign-in uses OAuth 2.1 with PKCE

Approval codes expire after 120 seconds and are tied to a secret only your app holds, so an intercepted code is useless. The consent screen shows exactly where your approval is sent.

Revoke in one click, effective at once

Every key is listed on your account page with when it was last used. Revoking one cuts it off on its next request. Reconnecting an app replaces its old key, and at most 10 app connections stay active.

Secrets are stripped first

Common API-key formats (OpenAI, Stripe, GitHub, AWS, Google, Slack), JWTs, database URLs and password=-style values become [REDACTED] before a prompt is processed and before anything is saved. It is pattern matching: it catches common formats, not every secret.

Rate limits cap the damage

A key is limited to 30 requests an hour and 100 a day on the free plan, with 20 written prompts a month, 120 an hour on Pro and 600 on Power. A leaked key can only write and check prompts, at a capped rate.

What PromptVibe stores

  • Your key’s hash and label — never the key itself.
  • For each request: which tool ran, the target tool you named, a task category, the framework, and any standing rules you stated (short lines, secrets stripped).
  • Your answers to PromptVibe’s follow-up questions, but only lasting preferences — audience, tone, format, style, how to check it is done, what it must not touch — as short lines with secrets stripped. One-off details, such as which file to edit, are never kept.
  • When you report back: accepted, edited or rejected, the lines you added, and a short note.
  • The personal skill built from those events, from your newest 1000 at most.

What it does not have

  • The text of your prompts, or of the prompts it writes, is not saved to your account.
  • Your files, repo, terminal or conversation. It is a remote service and only receives the text your agent puts in a tool call.
  • Your PromptVibe password. Apps get a connector-only key, never your login.

The trade-off to know about

To write or improve a prompt, the text your agent sends (after secrets are stripped) is passed to AI models from several providers, routed through OpenRouter, with a built-in template as the fallback. Those providers have their own data policies, which we do not control.

So do not put anything in a prompt that you would not want a model provider to see. diagnose_prompt, record_feedback and get_skill never call a model.

What your agent gets

Five tools. Only two call an AI model.

Every tool takes plain text, so any agent can use it. Name the target — “for Claude Code”, “for Midjourney” — and the prompt is shaped for it.

write_prompt

Turn a rough goal into one production-ready prompt for the tool you name — Claude Code, Cursor, Codex, ChatGPT, Midjourney and more. A vague goal gets a few questions first; the finished prompt is handed to your agent to run.

Calls an AI model

improve_prompt

Fix an existing prompt and see which known failure patterns it had.

Calls an AI model

diagnose_prompt

Instant check for the mistakes that waste the most time — vague task, no output format, no scope, no stop condition.

Read-only · no AI model

record_feedback

Tell it whether you accepted, edited or rejected a result. This is how your personal skill improves.

Saves your feedback · no AI model

get_skill

Download your personalised SKILL.md to use offline.

Read-only · no AI model

How it learns

It remembers what you keep correcting.

You use it

write_prompt and improve_prompt note the target tool, the kind of task, the framework and any standing requirement you state (“never touch the migrations folder”). When it asks a follow-up question, a lasting answer such as your audience or tone is noted too.

Your agent reports back

With record_feedback — accepted, edited or rejected. The lines you added and your reasons for rejecting a result are noted too.

A rule needs 2 prompts to stick

A line becomes a standing rule only after it shows up in 2 separate prompts, so one odd request never rewrites your skill.

It is deterministic counting, not a hidden model. Read exactly what it learned on your account page, download it as a SKILL.md, or erase it any time.

Keep a local SKILL.md in sync (optional)

Prefer an on-disk SKILL.md your agent loads by itself? This pulls your personalised skill; run it on a schedule or from a session-start hook.

# Refresh your personalised skill (run it on a schedule, or from a session-start hook)
curl -fsSL -H "Authorization: Bearer YOUR_KEY" https://www.promptvibe.co/api/mcp/skill \
  -o ~/.claude/skills/promptvibe-personal/SKILL.md --create-dirs

Free and Pro

Start free. Upgrade when it is part of your day.

Free

  • 20 written or improved prompts every month
  • Unlimited prompt checks and skill downloads
  • Every app and one-click setup on this page

Pro

  • No monthly limit on writing and improving prompts
  • 120 requests an hour per key (600 on Power)
  • Saved context, library, Compare Lab and scheduled tasks

The free allowance resets on the 1st of each month (UTC). When it runs out your agent tells you, and checks keep working. See Pro →

Help

Troubleshooting

401 — “Missing or invalid API key”

The key was mistyped, revoked, or sent in the wrong header. Keys start with pv_live_ and are shown once; create a new one from the setup panel above or on your account page. Check there is a space after “Bearer”.

429 — rate limit

The free plan allows 30 connector requests an hour and 100 a day, per key; Pro allows 120 an hour and Power 600. The response says when to retry. (If your agent says you have used all your free prompts, that is the monthly allowance, not a rate limit — see “Free and Pro” above.)

The tools do not appear

Restart the client after editing its config file, then check its MCP list (/mcp in Claude Code and Gemini CLI). Some clients need the server switched on per chat.

Claude or ChatGPT shows “couldn’t connect”

Use exactly https://www.promptvibe.co/api/mcp with no trailing path, and sign in with your PromptVibe account. If the sign-in window closes without finishing, try again once; if it keeps failing, use an API key with a client that supports headers and tell us.

Windows: the header is rejected

Quote the whole header value, or move it into an environment variable as in the bridge example; spaces in arguments are the usual cause.

“Unsupported transport” or SSE errors

The connector speaks Streamable HTTP, the current MCP standard. A client that only knows the older SSE transport needs the mcp-remote bridge.

Questions

Frequently asked

Is it safe to connect to my account?

The app gets a connector-only key, not your password, and it can only write, check and learn from prompts. Keys are stored as hashes, listed on your account page and revocable in one click, and secrets in a prompt are stripped before it is processed. The one trade-off — prompt text is processed by third-party AI models — is spelled out in the security section above.

Can it see my code or files?

No. It is a remote service that receives only what your agent puts in a tool call — the goal, prompt or context the agent chooses to send. It cannot browse your machine, repo or conversation.

Does it work with Claude?

Yes. Claude Code connects with one command, and Claude on the web, desktop and mobile can add it as a custom connector and sign in — no key to copy. Custom connectors are available on selected Claude plans.

Does it work with Codex?

Yes. Add the server to ~/.codex/config.toml with your key as an HTTP header; the setup panel creates the key for you.

Does it work with Grok?

Yes, through the xAI API: Grok supports remote MCP servers as a tool, and you pass your PromptVibe key in the tool’s headers. We could not confirm a custom-connector setting inside the Grok chat apps, so we do not claim one.

Does it work with ChatGPT, Gemini and Cursor?

Yes. ChatGPT supports custom MCP connectors with OAuth sign-in through developer mode; Gemini CLI and Cursor take a URL and an Authorization header (Cursor also has a one-click button).

What if my agent is not on the list?

If it supports remote MCP servers it should work: the URL, Streamable HTTP, and either a Bearer key or the OAuth sign-in. Clients that only speak the older SSE transport can use the mcp-remote bridge.

How is “works with” decided?

The protocol side is covered by automated tests that send each kind of client’s real messages, and every setup on this page follows the vendor’s own documentation or the standard configuration for that client (each is labelled). We have not run every vendor’s app ourselves, so if one misbehaves, tell us.

Will it ask me questions?

Only when it has to. If your request is short and vague (roughly ten words or fewer, with little context) and at least two details that matter are missing, write_prompt sends back up to three questions instead of a prompt. Your agent puts them to you in one message and calls again with your answers; say “skip” and it writes with sensible assumptions. It asks once per request, never asks what it already knows about you, and a round of questions does not use any of your monthly prompts.

Does it run the prompt automatically?

For write_prompt, yes by default: the result tells your agent to carry out the finished prompt straight away, so you do not copy, paste or confirm. That is an instruction to your agent, so your agent’s own permission prompts (before editing files or running commands, for example) still apply. It only makes sense when the prompt is meant for that agent: a Midjourney or ChatGPT prompt is handed to you ready to paste, because PromptVibe cannot send text into another app. Say “show me the prompt first” and the agent holds it for your approval. improve_prompt never runs what it returns.

Is it the same generator as the website?

Yes. The connector uses PromptVibe’s own prompt generator — the same rules and the same per-tool routes (v1.0.0, 25 routes) — and adds your personal preferences on top. The prompt is written by the same AI models the website uses, served through OpenRouter, which is why the security section asks you not to put secrets in prompts.

Is it really free?

Connecting is free, and so is the free plan: 20 written or improved prompts a month, with unlimited checks, at 30 requests an hour and 100 a day per key. Pro removes the monthly limit and raises the hourly one to 120; Power raises it to 600. Changes to these limits are posted on the changelog.

Does it cost AI credits?

No. Prompts are written on PromptVibe’s own model allowance and do not draw down your Compare Lab credits.

What happens if I cancel Pro?

Your keys keep working on the free plan (20 written prompts a month). Your learned skill is kept, and you can erase it any time.

Put your prompt writer where you work.

One-click setup, a connection you can revoke any time, and secrets stripped before anything is processed.