Key Concepts

The ideas that make the Dapta MCP click: two ways to authenticate, how the workspace is chosen, tools and guides, and the preview-then-confirm safety model.

A few ideas explain how the connection authenticates, how Claude knows what to do, and how your data stays safe.

---

1. Two ways to authenticate

• Sign in with your Dapta account. Recommended for people. Your AI client opens Dapta's sign-in page once (your email and a 6-digit code) and keeps a token. There is no key to copy. Claude, ChatGPT, Codex and Cursor all support it, and it is what the Claude connectors directory uses. Signing in reaches the workspaces where you are an owner or admin.

• API key. For scripts, automations, or a connection fixed to one workspace. The workspace owner creates keys in the Dapta admin panel. A workspace can have up to 10 keys, each with its own name, and each key is shown only once.

Treat an API key like a password. If it is ever exposed, revoke it and create a new one. Revoking a key does not affect your other keys.

---

2. Which workspace am I in?

• With an API key, the workspace is fixed by the key. Claude never asks.

• With a login, you may have several workspaces. Claude lists them by name and asks which one to use, then keeps using it for the conversation. Say "switch to <workspace>" to change.

Every answer starts with the workspace it ran in, so you always know where you are.

---

3. Tools and guides

• Tools are the raw actions: read a call, update an agent, patch a flow.

• Guides are the workflows that orchestrate tools for real tasks. A guide tells Claude the steps, the tool order and the points where it must confirm with you.

Claude loads a guide with the get_guidance tool when a task spans several tools. You never have to ask for it, and there is nothing to install: guides are served by the Dapta MCP server.

Older Dapta material called guides "skills". They are the same thing. There is no skills bundle to download anymore.

---

4. Preview, then confirm

This is the most important idea for trusting Claude with your workspace.

• Reads happen directly. Asking "how many calls did I get this week?" returns the answer. Nothing changes, so nothing needs confirming.

• Writes come in preview_* / commit_* pairs. The preview shows exactly what will change and returns a single-use token. Nothing is written until you say yes and the matching commit runs.

Three tools act without a preview: create_voice_agent creates a new agent and changes nothing that exists (Claude confirms the details with you first), discard_text_agent_draft drops an unpublished draft, and chat_with_text_agent runs a test conversation with a text agent.

An example

You say: "My agent talks too fast, slow it down."

1. Claude loads the guide for voice agent configuration with get_guidance.

2. It reads the agent with get_agent to see the current speed.

3. It calls preview_update_voice_agent with the new value. The server answers with a before and after comparison and a single-use preview token.

4. Claude shows you the comparison and asks for a yes.

5. On yes, it calls commit_update_voice_agent with the token. The server re-reads the agent and confirms what was written.

---

5. Tool hints

Every tool declares whether it only reads, adds data, or changes and removes existing data, and whether it can reach outside Dapta, for example by running a flow or queuing calls. Your AI client uses these hints to decide when to ask you before acting, and claude.ai groups the read-only tools separately.

---

Next step

Setup & Connection