Kipt MCP connects your private prompt and skill library to your coding agent. Once connected, you can ask Claude Code or Codex to find a prompt, fill its blanks, read a skill or save a useful new prompt without switching to the browser. Your GitHub repository remains the source of truth, and saving is a separate permission you choose to approve.
The hosted connection address is https://kipt.me/api/mcp. This is an authenticated agent endpoint, not a public listing of your library. Public guides and prompt collections remain separate from the private prompts and skills in your connected repository.
What can your agent do through Kipt MCP?
MCP means Model Context Protocol. It gives compatible clients a standard way to discover and call tools offered by another application. Kipt exposes tools for your library; the agent decides which to call for your request.
| Task | What happens |
|---|---|
| Find a prompt | Search returns matching prompts and skills with summaries and identifiers. The agent can then retrieve the relevant item. |
| Fill prompt variables | The agent supplies values for blanks such as {{product}} or {{audience:developers}} and receives the filled prompt. Missing values are reported. |
| Read a skill | The agent reads its instructions and supporting text files. This does not install the skill or run its scripts. |
| Save a new prompt | With saving approved, the agent creates a prompt with the inbox tag for your review. |
| Check the connection | Connection status reports the active authentication method, connected library, approved permissions and saved verification state. |
An MCP connection does not give an agent unrestricted control of your repository. Kipt does not expose tools to edit or delete existing items, create skill folders, install skills or publish projects.
Connect from Kipt Settings
Start with a connected library in Kipt. Open Settings, find Agent connection, then choose Connect MCP. Pick your agent and follow the setup instructions shown there. Those instructions are the right place to copy commands because they use the endpoint for the Kipt deployment you are visiting.
The browser sign-in flow asks you to review the requesting application and its access. Check that you are using the intended Kipt account and library. You can connect with read-only access, or check Allow saving new prompts before connecting when you want the agent to create prompts too.
After approval, return to your agent and restart or reconnect it so its tool list reflects the new connection. You can manage browser sign-in connections under Connected apps in the same Settings section. The general Codex MCP guide and Claude Code MCP guide explain the client-side concepts when you need more context.
Try finding and filling a prompt
Start with a prompt you already recognize. For example, ask: “Find my landing-page prompt in Kipt.” The agent can search, show matching entries and read the relevant prompt. A precise title or tag helps when several items cover the same subject.
If the prompt has variables, supply the values in your request: “Use my launch prompt with product set to Kipt and audience set to developers.” Kipt fills supplied values and uses defaults where available. If required blanks remain, the result identifies them so the agent can collect the missing information before using the prompt.
Retrieving a prompt does not change the stored template. You can reuse the same original with a different product or audience next time. Similarly, reading a skill gives your agent its instructions without automatically installing a folder on your computer. See skills versus MCP for how installed instructions and remote tools work together.
Save a useful prompt to your inbox
When you want to keep a new prompt, ask directly: “Save this to Kipt with the tag product-launch.” Saving needs approval on the connection the agent is actually using. Search access alone is not permission to create content.
New agent saves receive source: agent and the inbox tag, alongside any tags supplied with the request. Open Kipt and filter by inbox to review what was saved. You can edit or organize it in the app after that review. The agent's save tool cannot overwrite an existing prompt or remove something from the library.
If the same prompt text is saved again, Kipt reuses the existing item. That makes retrying a save less likely to leave duplicate copies. The underlying save is still a Git commit that updates the prompt and library index together, preserving the library's normal history.
Verify saving on the active connection
The Settings flow separates permission approval from proof that a save has completed. Verify saving gives you a request to copy into the connected agent. That request identifies the intended connection, checks its permissions and asks for an explicit save test.
The verification tool requires confirmation. It writes a disposable test prompt and a verification receipt, then compares the stored Markdown and index entry. Repeating the check reuses its test prompt while making a fresh verification write. After the agent finishes, refresh Settings to see the result.
A status response or a visible save tool is useful diagnostic information, but neither alone proves a new write succeeded. A past verified result also describes an earlier check; run verification again if you need to test current access after a change to your library connection.
If search works but saving does not
Check Connected apps first. If the active browser connection is read-only, choose Enable saving and complete a new browser approval with Allow saving new prompts checked. Restart or reconnect the client afterward so it discovers the newly available tools.
Manual tokens and browser connections have independent permissions. Creating a manual token that can save does not upgrade a client already signed in through OAuth. Equally, approving a browser connection does not change a separate manual token. A client configuration containing only the endpoint URL can already be using browser authentication.
For a longer walkthrough of this specific failure, read Codex MCP write access. Share only non-secret diagnostic details when asking for help: the client name, endpoint, active authentication method, approved permission and sanitized error. Keep access tokens and authorization headers out of chat.
Keep control of the connection
Review connected apps in Settings and revoke access when a client no longer needs your library. Revoking a browser connection does not revoke an independent manual token; manage each credential according to how it was created. Connecting a client also does not publish your prompts or add your private library to Kipt's public SEO pages.
Choose read-only access when your workflow only needs retrieval. Approve saving when you want the agent to capture useful work, then use the inbox tag as your review queue. The Kipt home page gives the short overview; the prompt library guide covers how to organize reusable templates once they are saved.