Connect an AI assistant (MCP)

The agaro MCP server lets Claude, ChatGPT, Cursor, and other compatible AI apps work in a workspace through Model Context Protocol. You connect an app by adding a connector URL and signing in with your Agaro account. No key is needed.

The hosted server calls agaro's authenticated API. It does not connect directly to the database. Workspace, module, role, model permission, read only checks, and the record and field controls implemented by the requested action remain in force. See Record scope and field masking for the supported coverage and release requirements.

#Where it lives

Open Settings, then the API and MCP tab, as a Super Admin. The tab has two blocks, in this order:

1. Connect an AI assistant (MCP), with the connector URL, the steps for each AI app, and the list of connected apps.
2. API keys, for your own code and systems that call the REST API. AI assistants do not need a key. See API overview and request format.

The tab needs the AI Assistant capability. Without it the tab says AI settings are not enabled and asks you to enable the AI Assistant capability first. Only a Super Admin is offered the tab, because API keys can carry owner authority.

#Requirements

• An active, paid Scale or Enterprise workspace. MCP is included with the Open API at no extra cost: there is no seat to buy or assign. A trial, the Free plan, and Essentials do not include it.
• The AI Assistant capability.
• A person whose account can connect apps. The person who adds the connector signs in and approves it, and the app then acts as that person.
• An AI app that supports remote MCP servers with OAuth.

Agaro checks these on the approval page. If one is missing, the page says why and nothing is connected. The wording differs by case, for example that the workspace needs a paid plan with API access, that its subscription is inactive, or that its plan does not include API and MCP access.

#Connect an AI app

The connector URL is https://mcp.agaro.ai/mcp. Copy it from the Connector URL field on Settings → API and MCP, which is read only and has a Copy button, so you always paste the exact address your workspace shows. Then choose your AI app in the row of choices named Choose your AI app.

#Claude

1. In Claude on the web or the desktop app, open Settings, then Connectors.
2. Choose Add custom connector.
3. Paste the connector URL, then choose Connect.
4. Sign in with your Agaro account and approve the connection.

#ChatGPT

1. In ChatGPT settings, add a custom or remote MCP connector. Where ChatGPT offers this depends on your plan.
2. Paste the connector URL. If it asks how to sign in, choose OAuth.
3. Sign in with your Agaro account and approve the connection.

#Cursor

1. Open your Cursor mcp.json file.
2. Add the server below, then save the file. The tab shows this block with its own Copy button and the connector URL already filled in.
3. When Cursor offers to connect the server, sign in with your Agaro account and approve the connection.

{
  "mcpServers": {
    "agaro": {
      "url": "https://mcp.agaro.ai/mcp"
    }
  }
}

#Other MCP clients

1. Use any AI app or client that supports remote MCP servers with OAuth.
2. Add a custom or remote MCP connector with the connector URL. If it asks for a transport, choose Streamable HTTP.
3. Sign in with your Agaro account when the app opens the sign in page, then approve the connection.

#Approve the connection

The approval page is titled Allow followed by the app name, then to use Agaro ERP? It shows the app, where approving sends you back to, who you are signed in as, and the workspace. It says the app will work in your workspace as you, with exactly the access you have, and lists what it can do:

• Read and change the records, files, and settings you can open in Agaro.
• Use only the modules and permissions you have. It cannot do anything you cannot do yourself.
• Keep this access until you disconnect it, or until it goes unused for 30 days.

Approve only if you started connecting the app yourself. Select Approve to connect or Cancel to stop. An app whose maker Agaro cannot confirm is marked unverified, with a warning that anyone can give an app any name. Approve it only if you recognize the address it returns to.

#Connected apps

An app acts as the person who approved it, with that person's access. It never acts as a shared administrator.

• Super Admins see every connected app in the workspace under Connected apps on the API and MCP tab. The table shows the app, the person, when it was connected, and when it was last used, with Disconnect on each row.
• Everyone sees their own connected apps under Account Settings → Security and Privacy → Connected apps, and can disconnect them there.

Select Disconnect and confirm. The app stops working with Agaro at once, and it can be connected again later. A Super Admin can disconnect any person's app. Disconnecting is allowed even while a workspace is read only.

#Test safely

Begin with erp_connection_context to identify the authenticated workspace and actor, then use discovery to find an appropriate read action.

Then verify:

1. The result comes from the intended workspace.
2. A record you cannot access is also unavailable through the connected app.
3. A low risk test write follows the expected model and role permission.

Do not begin with a send, refund, delete, charge, or payroll mutation.

#Authority model

The connected app acts as the person who approved it. Their role, module access, record scope, and field visibility apply to every call. Ask a person with the access the job needs to connect the app, instead of connecting it as a more powerful person.

MCP is separate from the in app assistant autonomy toggle. External MCP calls do not become safer because the chat composer is set to Action on Accept. Your AI app is responsible for its own confirmation experience, while agaro enforces the signed in person's access and action permissions.

Field masking is enforced by the ERP on supported employee responses, not by a visual mask in the AI app. A missing field stays absent from both the text and structured result. Test with the intended restricted person rather than using a Super Admin as a substitute.

#Tools and action coverage

The hosted bridge exposes MCP tools that dispatch supported registered actions. Named tools provide convenient common operations. The underlying authenticated action surface determines the product capability available to the bridge.

Tool names and action coverage can change as the product evolves. Let the AI app refresh tool discovery instead of hardcoding a static catalog into instructions.

Unavailable module actions and explicit model denials are omitted from supported discovery. Named calls refresh availability, and the ERP still checks authorization at execution. A stale tool list is not a permission grant. Writes are not automatically retried after an uncertain response; inspect the recorded outcome first.

#Troubleshooting

#The approval page says the workspace needs a paid plan

MCP needs a paid Scale or Enterprise workspace. A trial, the Free plan, or a read only workspace is refused. A Super Admin can subscribe or restore billing under Settings → Subscription, then connect the app again.

#The approval page says the plan does not include API and MCP access

MCP is included in Scale and Enterprise, not in Essentials. Change the base plan to Scale or Enterprise.

#The approval page says you are working in another workspace

Switch back to your own workspace, then connect the app again. A connection always belongs to your own workspace.

#The approval page says your account cannot connect apps

Some kinds of account cannot connect apps, and the page says This kind of account cannot connect apps to Agaro ERP. Use a full user account.

#The app cannot connect or says the link is damaged

Start connecting again from the AI app. A connection link that is incomplete, or that points at an address the app did not register, is stopped for your safety and nothing is connected.

#The app stopped working

A connection ends when someone disconnects it or when it goes unused for 30 days. Connect it again from the AI app.

#Calls return 429

Reduce request concurrency and wait at least the returned retry interval. MCP tool loops can produce several API calls from one natural language request.

#The app can read but cannot write

Check the person's role, model permissions, record scope, installed module, paid entitlement, and workspace subscription state. MCP does not bypass any of them.