Connectors through the MCP server

A connector is Python code that runs in the Nuntiq integration worker and moves data between Nuntiq and an ERP or other system. The Connector Framework section describes the runtime and the local toolkit. Through the MCP server you can set up, write and test a connector on a customer without leaving your AI app.

Recommended way of working

Build the connector locally first, with the Nuntiq Connector Toolkit and your AI agent, and move it to the customer only when it works.

1. Develop locally. Download the toolkit and open it in your AI agent (Claude Code, Cursor, VS Code). The toolkit ships an AGENTS.md (the guide for coding assistants) that gives the agent the runtime contract, the entrypoint keys and the Nuntiq API. Let the agent write the connector in files.
2. Test locally. Run the connector with the toolkit. Nuntiq is simulated with dummy data, so nothing reaches a customer. With credentials for a test or sandbox system of the ERP in the local settings file, the entrypoints call that real system, so you test the integration itself before anything is configured.
3. Move it to the customer. Ask the AI app to propose the connector, its code, integration rules and scheduler tasks through the MCP server as a change set. Read the diff and approve.
4. Enter the secrets in the admin portal.
5. Test on the customer. Run a scheduler task now, read the run history and what the connector printed, and fix through a new change set.

Fixing on the customer is slower than fixing locally: every round is a change set, an approval and a real run. Keep the local copy as the source of truth, make larger changes there first, and propose them again.

What the AI app can do on the customer

• Pick a connector template (get_catalog with templates, kind connector) and propose the connector from it.
• Write and change the connector code (entrypoints and shared modules) in a change set. The code is in the diff for you to read.
• Propose the integration rules (which invoices go to the ERP, and when) and the scheduler tasks for master data runs.
• Run a scheduler task now (trigger_scheduler_task). This is a real run; the app asks you first.
• Read the run history (get_scheduler_history) and what the connector printed (get_task_output), then correct the code, settings or rules through a new change set and run again.

What stays with you

• Secrets. API keys and passwords of the external system are never set through the AI app. The app tells you which secret keys to fill in the admin portal (Connectors, the connector, Secrets).
• Reading the code. The code runs with the connector's secrets. Before you approve a change set with connector code, check in the diff that it never prints a secret or sends one anywhere but the external system.

Connector output read through the MCP server has secret values masked as [***]. You see the same output in the portal under View output in the scheduler history, and under View connector output on an invoice's Process tab.

Toolkit and MCP server

	Connector Toolkit	MCP server
Where the code runs	Your machine; Nuntiq is simulated with dummy data	The customer's environment, in the integration worker
Who writes the code	Your AI agent, in files	Your AI app, in a change set
Secrets	Your local settings file (use test credentials)	Entered in the admin portal
Use it for	Writing and testing the connector, debugging step by step, version control	Putting the connector on the customer, small corrections, checking a real run

Both work on the same connector: code from the toolkit is proposed through the MCP server unchanged, and code read through the MCP server can be taken back into the toolkit for the next round.