Talkward × MCP-Server
Your inbox for every AI agent
The Talkward MCP server gives AI agents access to your inbox through the open Model Context Protocol. Claude Code, Claude Desktop, Cowork, ChatGPT or your own agent search conversations, read histories, write notes and set tags. Every agent with its own key, revocable at any time.
Sound familiar? Your AI agent can do a lot, except look into your inbox.
You work with Claude or ChatGPT every day. But for every customer service question you copy conversations back and forth.
Copy, paste, ask.
To summarise ten complaints you paste ten transcripts into the chat. Then you copy the answer back.
A custom integration for every agent.
Your developer builds an agent that pre-sorts requests. First he writes an API client, with sign-in, paging and error cases.
Nobody sees what the AI did.
A script sets tags and writes notes. Who did it and when is recorded nowhere.
With the MCP server you connect an agent in a minute. Everything it changes shows in the conversation with its name.
ToolsTen tools for your agent.
These are all the tools of the server. Reading tools work with any key, writing ones only if the key allows it.
| Tool | What it does | Kind |
|---|---|---|
| gespraeche_suchen | Search conversations | reads |
| gespraech_lesen | Read a conversation with its history | reads |
| kontakte_suchen | Search contacts | reads |
| kontakt_lesen | Read a contact | reads |
| auftraege_offen | Fetch open AI jobs | reads |
| notiz_schreiben | Write an internal note | writes |
| antwort_senden | Send a reply to the customer, not with “read only” | writes |
| status_setzen | Set status | writes |
| tags_setzen | Set tags | writes |
| auftrag_abschliessen | Complete a job, report the result | writes |
Set upOne line or one file.
In Claude (web, desktop, Cowork) Talkward is in the connector directory: search for “Talkward”, connect, sign in with Talkward. In ChatGPT you enter the URL as a connector. In Claude Code a command or a file in the project with a key is enough.
claude mcp add --transport http talkward https://chat.talkward.app/api/mcp \
--header "Authorization: Bearer tw_…"{
"mcpServers": {
"talkward": {
"type": "http",
"url": "https://chat.talkward.app/api/mcp",
"headers": { "Authorization": "Bearer tw_…" }
}
}
}What agents do with it
Four ways to work your inbox with AI.
The same server for questions in chat, jobs from automations and your own agents.
Questions to your inbox
In Claude or ChatGPT: “Summarise all open complaints from today.” The agent searches and reads on its own.
Drafts for approval
“Reply to the three oldest conversations in our brand voice, as a note for approval.” Nothing is sent until your team wants it.
Jobs from automations
The block “Claude Code beauftragen” creates a job. The agent reads the conversation, works and reports back with auftrag_abschliessen.
Your own agents
Your agent pre-sorts, sets status and tags or prepares replies, without its own API client.
How it worksOne server, one key per agent.
The MCP server runs at https://chat.talkward.app/api/mcp and speaks the Model Context Protocol over HTTP. An agent signs in with a Talkward API key and gets exactly what the key allows. Every change shows in the conversation with the key’s name.
One key per agent
Every agent gets its own Talkward API key. Its name shows next to everything it does in the conversation.
“Read only” and “Notes only”
“Read only” searches and reads. “Notes only” also writes notes, tags and results, but never to customers.
Revocable at any time
Revoke the key and the agent is out immediately.
Connected to automations
The agent fetches jobs from “Claude Code beauftragen” with auftraege_offen and completes them with auftrag_abschliessen.
- Connection
- Model Context Protocol over HTTP at https://chat.talkward.app/api/mcp, bearer key per agent
- Access
- Ten tools: search and read conversations and contacts, open jobs, note, reply, status, tags, complete job
- Processing
- Talkward: Nuremberg (ISO 27001). What an agent reads is processed by its provider; for Claude and ChatGPT in the USA, under your own DPA.
What an agent leaves in the conversation.
Notes, status and tags carry the key’s name. Your team sees which agent did what and can take its access away at any time.
- Key per agent
- “Read only”
- Name in the conversation
- Ten tools
- Revocable any time
- Open jobs
- Rights
- Read and write
- Note
- Double delivery caused by two active subscriptions
- Tag
- Double subscription
- Sender
- Claude Code · Support agent
- Revocation
- any time, effective immediately
Set up in four steps.
The MCP server is included from Professional at no extra cost.
Switch on the add-on
Under Settings › Add-ons switch on the matching add-on: Claude, ChatGPT or the API for your own agents.
Choose the rights
With “Sign in with Talkward” or a key per agent: “Read only”, “Notes only” or “Read and write”. Every connection can be disconnected under Add-ons.
Add the connector
In Claude Desktop, Cowork or ChatGPT as a connector with the URL https://chat.talkward.app/api/mcp, sign-in via OAuth. In Claude Code by command with a key as bearer token.
Get going
Ask the agent about your inbox or drag “Claude Code beauftragen” into an automation.
MCP server
no extra costfrom Professional, provisional, net
No extra cost from Professional. The models of your agents are billed by their providers through your account.
All plans and prices- Ten tools for conversations, contacts and jobs
- Model Context Protocol over HTTP
- Key per agent, “read only” possible
- Every change in the conversation with its sender
- Connected to “Claude Code beauftragen”
- Servers in Nuremberg (ISO 27001)
- Data processing agreement
- AI replies labelled (EU AI Act)
- Extensions cancellable monthly
Questions about the MCP server
Question not listed? Write to us at hey@talkward.app.
Which AI agents work?
Any that speaks the Model Context Protocol over HTTP: Claude Code, Claude Desktop and Cowork, ChatGPT with connectors in developer mode, and your own agents.
How does an agent sign in?
With a Talkward API key as a bearer token. Create one key per agent, then its name shows next to everything it does in the conversation.
Can I limit access?
Yes. A “read only” key searches and reads but sends no replies to customers. Every key can be revoked at any time, and the agent is out immediately.
What does the MCP server cost?
Nothing extra, included from Professional. Claude, ChatGPT or other models are billed by their provider through your own account.
Where is the data processed?
The MCP server is part of Talkward, storage stays in Nuremberg. What an agent reads is processed by its provider; Claude and ChatGPT are services in the USA with which you sign your own data processing agreement.
What are auftraege_offen and auftrag_abschliessen for?
For jobs from automations: the block “Claude Code beauftragen” creates a job, the agent fetches open jobs and reports result and outcome with auftrag_abschliessen. The automation then continues.
Goes well with
Claude integrationBrief Claude Code from an automation: fix bugs, research, prepare quotes. The result comes back.
ChatGPT integrationChatGPT as a block: extract inquiries, rate mood, drafts in your brand voice. With your OpenAI key.
APIConnect your own systems: read conversations, send messages, set fields, webhooks. Documented, your own keys.- All integrationsShop, ERP, marketplaces, shipping and service
Secureyour seat.
Three details, no commitment. We will get back to you personally within two working days.