MCP Server

Connect an external AI coding tool — Claude Code, Codex or Antigravity — directly to your AiTalk account. Once connected, you describe what you want in plain language and the AI builds it: agents, workflows, knowledge base files.

What It Is

MCP (Model Context Protocol) is a standard way for an AI tool to use an outside service. AiTalk acts as the MCP server; your AI tool is the client.

Your AI tool (Claude Code / Codex / Antigravity)
        │
        │  Authorization: Bearer aitk_...
        ▼
AiTalk MCP Server  ──→  your agents, workflows, knowledge base, conversations

What the connected AI can do:

Area	What it can do
Agents & workflows	List them, read them, create new ones from templates or from scratch, edit them, validate them
Deploy	Take a workflow live, or switch a live workflow back to draft
Knowledge base (RAG)	List, read, add and delete source files
Conversation history	Read past text chats, voice calls and quiz attempts

What it cannot do:

Delete workflows or agents — use the dashboard
Run a test — use the dashboard
Change your subscription, billing or credits
Reach any other customer's data. Every request is limited to the account that owns the token

Why You Need an Access Token

The access token is how your AI tool proves it is you. There is no other way in — without a token, every request to the MCP server is rejected. Nothing can connect until you create the first one.

A few things worth knowing:

The token is shown only once, right after you create it. Copy it immediately. If you lose it, revoke it and create a new one.
Write access is on by default, so the AI can build and edit. Uncheck it if you want a read-only token (useful when you only want the AI to look, not touch).
You can have up to 5 tokens. One per machine or per tool is a good habit — it keeps them easy to revoke individually.
Treat a token like a password. Anyone holding it can act on your account as you.

Setup

Step 1: Create a Token

In AiTalk, go to Settings → MCP Server
Type a name — the machine or tool you will use it on, e.g. Claude Code - MacBook
Leave Allow write tools checked (uncheck it for a read-only token)
Click Create token
Copy the token now — it is shown only once

Step 2: Add AiTalk to Your AI Tool

Under Connect your AI tool, pick your tool's tab. The snippet already contains your new token — copy it and use it as described below.

Claude Code — run the command in your terminal:

claude mcp add --transport http aitalk https://www.aitalk.ch/api/mcp-server \
  --header "Authorization: Bearer aitk_YOUR_TOKEN"

Claude Desktop — Settings → Developer → Edit config, paste the snippet, then restart the app. Requires Node.js (npx). Config file location:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

Codex — add the snippet to ~/.codex/config.toml. In the Codex app: New MCP → Type = Streamable HTTP, then enter the URL and an Authorization: Bearer header.

Antigravity — add the snippet to ~/.gemini/config/mcp_config.json (Settings → Customizations → Open MCP Config), then refresh.

Step 3: Ask Your AI

Just talk to it. For example:

"List my AiTalk agents and workflows"
"Show me what the customer support workflow does"
"Create a workflow from the FAQ template and set the greeting to ..."
"Add this text to my knowledge base"
"Deploy the booking workflow"

The AI reads AiTalk's built-in guides on its own, so you do not need to explain how workflows are structured.

Monitoring Access

Two places show you how your tokens are being used:

Settings → MCP Server — every token, with when it was last used and the IP address and country it was used from
Dashboard — a summary: how many tokens you have, and when and where the most recent access came from

If you see an access you do not recognise, revoke that token immediately (the trash icon next to it). Connected tools lose access at once.

The country is resolved from the IP address of the connection. A VPN or a cloud-hosted AI tool will show the country of that server, not yours.

Safety Rails

A few limits exist so that a mistake — or an AI that misreads an instruction — cannot do quiet damage:

Deploying is visible. Taking a workflow live is a deliberate action. If you would rather review before anything goes live, ask your AI to request a deploy instead: it returns a link that you open and click Deploy on yourself.
Editing a live workflow reaches your customers immediately. Your AI is told to confirm with you first.
Secrets are never handed out. API keys, passwords and tokens inside your workflows are masked before anything leaves the server, so a connected AI cannot read them.
Rate limits apply per token (reads and writes are limited separately).
Every call is logged — which tool ran, whether it succeeded, and how long it took.

Troubleshooting

Problem	Cause / Fix
`401 Unauthorized`	The token is wrong, was revoked, or the `Authorization: Bearer` prefix is missing. Create a new token and re-add it.
Tool connects but shows no agents	The account holding the token has no agents yet. Create one in the dashboard first.
Write actions rejected	The token was created read-only. Create a new token with Allow write tools checked.
"Rate limit exceeded"	Too many calls in one minute. Wait a moment and continue.
A new source file is not found yet	Indexing runs in the background. Wait a few seconds and ask again.
Cannot upload a PDF / Word / Excel file	Binary files cannot travel over MCP. Have your AI convert the content to text first, or upload the file on the dashboard Storage page.