MCP Server
Build and run your AI agents from Claude Code, Cursor or VS Code
ThinnestAI is an MCP server. Connect your coding assistant with one API key and it can build voice and WhatsApp agents, connect them to your own database, test them, run your inbox and build your dashboards — and nothing reaches a customer without your approval.
In short
What can a coding assistant do with ThinnestAI?
Through the MCP server, Claude Code, Cursor or VS Code can build your voice and WhatsApp agents, connect them to your API, test-chat them, read conversations and calls, and run your inbox and campaigns. Anything that would message a customer or place a call waits for your approval.
Plain words
Ask for it in plain words
Your assistant works through the ThinnestAI API, so it can do exactly what its key can do — and nothing more.
01
“Create a support agent for our clinic that answers in English and Spanish, and teach it from our website.”
Creates the agent, writes its instructions, adds your pages as knowledge, then test-chats it.
02
“Connect it to our appointments API so it can check open slots.”
Adds an action pointing at your endpoint, tests it, and asks you before switching it on.
03
“What did customers ask this week that the agent couldn't answer?”
Reads the agent's analytics, groups the questions, and drafts knowledge for you to approve.
04
“Go through the conversations waiting for my team.”
Summarises each one, suggests reply, assign, take over or resolve — and does what you decide.
05
“Pause the Diwali broadcast — the offer code is wrong.”
Pauses the campaign. Messages already handed to Meta still complete; resuming asks for your approval.
06
“Build me a dashboard of yesterday's calls and what they cost.”
Writes the code against the API with a read-only key and sets up the webhooks to keep it live.
Connect
Connect in one line
Create a key in Settings → API keys. The console shows each client's config ready to copy, under Connect an AI assistant.
Claude Code
claude mcp add --transport http thinnestai \ https://app.thinnest.ai/api/v1/mcp \ --header "Authorization: Bearer ta_live_YOUR_KEY"
Cursor (~/.cursor/mcp.json)
{
"mcpServers": {
"thinnestai": {
"url": "https://app.thinnest.ai/api/v1/mcp",
"headers": { "Authorization": "Bearer ta_live_YOUR_KEY" }
}
}
}VS Code, Claude Desktop and Windsurf: see the connect guide. Then ask: “List my ThinnestAI agents.”
Keys
Choose what the key can do
The MCP server only shows the assistant the tools its key may use. Start with a Build key.
01 · Recommended
Build
Build agents, connect your API, manage contacts and webhooks, resolve and assign conversations, pause and cancel campaigns, read everything, test-chat.
Message a customer, send a code or place a call — those tools are not even offered.
02 · Dashboards & reports
Read-only
Read conversations, calls, contacts, analytics and settings.
Change anything.
03 · When you want it to send
Full access
Everything, including sending, calling and resuming campaigns.
Do any of those without your approval in the conversation.
Safety
Safe to hand to an assistant
Customer messages are treated as information, never as instructions — and the checks hold even when the assistant is wrong.
Nothing reaches a customer without your yes
Tools that message, call, delete or decide where customer data is sent are marked destructive, and refuse to run until the assistant confirms you approved — even in a client set to auto-approve.
The key is the boundary
The assistant can do exactly what its API key can do over the API — same checks, same limits. Revoke the key and its access ends immediately.
Nobody is contacted twice
Sending and calling tools need a retry key, so an assistant that retries after a timeout gets the first answer back instead of a second message.
See what an assistant did
Settings → API keys lists the changes assistants made through your keys, newest first, and every time one was stopped to ask — the tool and the outcome, never what it was asked.
Test chat
Test the agent without leaving the editor
test_chat talks to your real agent — same instructions, model, knowledge and actions — and returns its reply with every tool it used and what went in and out.
Anything that would phone, message, email or send a code to a person, or change a booking, is switched off. The agent says what it would have done instead, so you see the behaviour and nobody is contacted.
Your own database
Keep your records in your own database
Instead of uploading CSVs, give the agent small endpoints of yours — “find this order”, “check this slot”. It asks for one answer at a time, mid-conversation, and your records stay where they are. Leads, escalations and call results come back to you by webhook.
What we do keep: the conversation history, so the agent can continue it — and what each action was called with, as an audit trail. What stays where
Tools
More than 150 tools, one API key
Everything the console does — grouped the way you think about your workspace. Anything that needs a secret, like a CRM key, stays in the console.
Agents
List, create, update and delete agents; models and voices; test chat.
Knowledge
Add website pages and text an agent answers from.
Actions
Connect your own HTTPS endpoints — create, test, enable.
Contacts
Look up, add and update contacts; record events.
Conversations
Read messages, filter by status; resolve, hand back, take over, assign (paid plans).
Templates, sequences, template sends and verification codes.
Calls
Place one call or a batch, read transcripts, cancel scheduled calls.
Webhooks
Create endpoints, send a test event, read the delivery history.
Campaigns
Draft and launch (with approval), results, pause, resume, cancel.
Usage & analytics
Usage and balance, agent analytics, call totals, WhatsApp cost by category.
Templates & forms
Draft WhatsApp templates and forms, send them for review, publish.
Sequences
Build drip sequences step by step; switch them on with your approval.
Phone numbers
Search, rent and release numbers with approval; point them at agents.
Team
Projects, teams, roles, invitations and saved replies.
Billing
Daily usage, the call log, the wallet ledger and invoices.
Prompts
Ready-made prompts
Type / in Claude Code to see them. Each is a worked plan, with the points where you approve written in.
Honest fit
What works today
Works now
- —Claude Code, Cursor and VS Code — directly
- —Claude Desktop, Windsurf and other local-only clients — through mcp-remote
- —Any MCP client that can send an Authorization header
Not yet
- —Connectors inside claude.ai and ChatGPT — they sign in with OAuth, which is not available yet
- —Anything that needs a secret — connecting a CRM, Cal.com or a carrier account with its key — is done in the console, so no key passes through your assistant
Read more
Read more
FAQ
MCP server FAQs
What is the ThinnestAI MCP server?
+
A Model Context Protocol server at https://app.thinnest.ai/api/v1/mcp, the same address for every workspace. Connect an AI assistant such as Claude Code, Cursor or VS Code to it with a ThinnestAI API key, and the assistant can build and run your voice and WhatsApp agents through the same API your own code would use.
Which AI assistants work with it?
+
Claude Code, Cursor and VS Code connect directly. Claude Desktop, Windsurf and other clients that only run local servers connect through the mcp-remote bridge. Connectors inside claude.ai and ChatGPT sign in with OAuth, which is not available yet.
Can the assistant message my customers without asking?
+
No. Every tool that messages a customer, places a call or deletes something refuses to run until you approve it in the conversation. With a Build or Read-only key, the sending and calling tools are not available at all.
Does my data stay in my own database?
+
Your records do. With actions, the agent asks your API for one answer at a time instead of you uploading your data. The conversation itself passes through ThinnestAI and the AI model, and we keep the conversation history so the agent can continue it.
Can the assistant create and launch a WhatsApp broadcast?
+
Yes, with your approval. The assistant builds the campaign as a draft — audience, template, blanks — tells you how many people it reaches, and launches it only after you say yes. Launching needs a full-access key. Scheduling a start time is done in the console.
Does it cost extra?
+
No separate charge for the MCP server. What the assistant does is billed as it would be from the console or the API — calls, messages, test chats and knowledge indexing use your plan as usual.
Connect your coding assistant with one API key
Build voice and WhatsApp agents, test them, run your inbox and build your dashboards — and nothing reaches a customer without your approval.