For AI Agents
Connect Your AI to Reply.la
Connect ChatGPT, Claude, or Claude Code to your Reply.la account with MCP so your AI can triage your Inbox, manage contacts, and build chatbot drafts for you.
What This Is
MCP for AI Agents lets you connect an AI app like ChatGPT or Claude to your Reply.la account. Once connected, you can talk to your AI in plain language and it works inside your account for you:
- "What needs my reply first today?"
- "Why didn't my chatbot reply to Aisyah?"
- "Add our new Raya menu to the chatbot."
Your AI reads your Inbox, contacts, chatbots, and Analytics, and can make changes you approve. It only ever sees the one organization you connected.
This is different from the documentation MCP server on For AI Agents. That one only lets an AI read these help docs. This one connects to your actual Reply.la account.
Once connected, see How to Use MCP for example questions, sample answers, and the permission each feature needs.
Before You Start
- You need a Reply.la plan that includes MCP for AI Agents. If you don't see it under Integrations, your plan doesn't include it yet, or your organization isn't in the beta yet. Contact Reply.la support to ask about access.
- You need an AI app that supports MCP: ChatGPT, Claude, Claude Code, or any other app with MCP connectors.
- Only the organization owner can connect, unless your role has been given permission to connect AI agents.
- A connection belongs to you. It works while you're a member of the organization and stops when you leave. Your AI can never do more than your own team role allows.
Server Address
All apps connect to the same address:
https://dash.reply.la/api/public/v1/mcp
You can also copy it from the dashboard: go to Integrations → MCP for AI Agents → Connect.
You never need an API key or token. Your AI app opens a Reply.la sign-in and approval screen the first time it connects.
Set Up Your AI App
Claude (web or desktop)
- In Claude, open Customize (or Settings), then choose Connectors.
- Click + Add, choose Add custom connector, give it a name, paste the server address, and click Continue.
- Keep the other settings as they are and click Add. Claude registers itself, so you don't need a client ID, secret, or request headers.
- The first time Claude uses Reply.la, it opens an approval screen. Choose your permissions and approve it there.
According to Anthropic's docs, custom connectors work in Claude, Cowork, and Claude Desktop on the Free, Pro, Max, Team, and Enterprise plans. The Free plan allows one custom connector. On Team and Enterprise, an owner adds the connector first in Organization settings → Connectors, then each member clicks Connect.
ChatGPT desktop app
- In the ChatGPT desktop app, open Settings, then choose Plugins.
- Choose Add, then Add MCP server.
- Give it a name such as
Replyla, set Type to Streamable HTTP, and paste the server address into URL. - Leave the bearer token and headers empty, then press Save. Reply.la signs you in instead.
- Choose Authenticate next to Replyla and approve on the Reply.la screen. If it's not there yet, restart ChatGPT.
ChatGPT on the web
- In ChatGPT in your browser, open Settings, then choose Plugins.
- Choose Add, then Add MCP server.
- Give it a name such as
Replyla, set Type to Streamable HTTP, and paste the server address into URL. - Leave the bearer token and headers empty, then press Save.
- The first time ChatGPT uses Reply.la, it opens an approval screen. Approve it there.
Switch On Developer Mode in ChatGPT
If there's no Add button under Plugins, switch on Developer mode first:
- Open Settings in ChatGPT on the web at chatgpt.com. Developer mode isn't available on mobile.
- Find Developer mode under Security and login. If it isn't there, look in Apps → Advanced settings.
- Switch the toggle on. No toggle at all? Your account hasn't been given access yet (see the note below).
- Go back to Plugins. The Add button now appears, so continue from step 2 above.
According to OpenAI's docs, full MCP support including write actions is in beta for ChatGPT Business, Enterprise, and Edu. On Business, only admins and owners can switch on Developer mode. On Enterprise and Edu, an admin gives access first in Permissions & Roles. On Plus and Pro, write actions may be limited. ChatGPT may also ask you to confirm before it runs a write action.
Claude Code
- Run this command in your terminal, from any folder:
claude mcp add --transport http replyla https://dash.reply.la/api/public/v1/mcp - Inside Claude Code, type
/mcp, choose replyla, then choose Authenticate. - Your browser opens the Reply.la approval screen. Choose your permissions and approve it there.
- Run the command in your normal terminal, not inside a Claude Code session.
- Before you sign in,
claude mcp listshows Needs authentication. That's normal. - By default the server is only added to the current project. Add
--scope userto the command to use it in all your projects.
Other MCP apps
- Open your app's connector, integration, or MCP server settings.
- Paste the server address. Most apps register themselves, so a client ID and secret are usually not needed.
- If your app does ask for a client ID, go to Integrations → MCP for AI Agents → Connect, choose My app is not listed, it supports MCP, and register a connector under My app is asking for a client ID. Give your app that ID along with the server address.
- Approve on the Reply.la screen the first time your app connects.
Choose What Your AI Can Do
On the approval screen, pick a preset or tick each permission yourself:
- View only (recommended, and the default) - Reads your data. Changes nothing.
- View and edit - Edits contacts and the Inbox, creates chatbots, and saves drafts. Never messages customers.
- Full access - Also sends messages, deletes contacts, and tests, trains, and publishes chatbots.
- Custom - Choose each of the 20 permissions yourself.
See How to Use MCP for the full permission map and what each feature needs. Your AI only sees the tools you approved, and only if your plan includes that feature.
Safety Built In
Reply.la gives every AI the same rules, so your customers stay safe:
- Chatbot changes are drafts. Nothing your AI changes in a chatbot reaches customers until it's published. Your AI will tell you "a draft is waiting", not "your chatbot is updated". The one exception: if an Auto Label change names a label you don't have yet, that label is created straight away, so your AI asks you first.
- Publishing needs your yes. Before publishing, your AI shows you a summary of every change and any warnings, then asks you plainly. The confirmation expires after 10 minutes, or as soon as the draft changes.
- Training needs your yes. Training AI Knowledge uses AI tokens, so your AI shows you what it will train and asks first.
- Broadcasts are never sent. Your AI can only create draft campaigns. A person reviews and sends them from the dashboard.
- Messages are confirmed first. Your AI confirms who it's messaging and what it will say before sending a real WhatsApp message.
- The playground is shared. The chatbot playground your AI uses is the same one you see in the dashboard. Test messages use AI tokens, and your AI asks before clearing it.
- Customer text is treated as data. If a contact name, message, or note contains something that looks like an instruction, your AI ignores it and warns you.
Reply.la tells your AI to ask you before publishing, training, or sending, but it can't check your answer. Only connect AI apps you trust, and only tick the permissions you need.
Manage or Disconnect
Go to Integrations → MCP for AI Agents → Connected agents to see every AI app connected to your account. From there you can:
- Click Edit access to change what a connection is allowed to do.
- Click Revoke to remove a connection. The AI app stops working with Reply.la immediately.
You can only change your own connections, unless your role lets you manage everyone's.
Common Questions
No. A connection is bound to one organization. No tool accepts an organization ID, so your AI can't reach another organization's data.
No. Every chatbot edit goes into a draft. Customers only see it after you publish, either from the dashboard or by saying yes when your AI shows you the publish summary.
No. Your AI can only create a draft campaign. Sending is always done by a person in the dashboard.
The tool list is filtered for each connection. A tool only appears if you approved its permission, your plan includes the feature, and your organization has access to it if it's still in beta. Check Edit access under Connected agents.
Reading data doesn't. Testing in the chatbot playground and training AI Knowledge do use your organization's AI tokens, the same as doing it in the dashboard.
Their connections stop working when they leave. For an automation that must keep running regardless of who's on the team, use an organization API token with the Reply.la API instead. Note that an API token works with the API only, not with AI apps through MCP, and it covers contacts, templates, messaging, and devices.
For AI Agents
Use Reply.la's documentation with ChatGPT, Claude, Perplexity, and other AI tools. Includes ready-to-paste links and a public MCP server.
How to Use MCP
What you can ask your AI agent once it's connected to Reply.la, with example questions, sample answers, and the permission each one needs.