On this page
Build with an AI agent
A coding assistant can help you build and maintain a Cumulus agent inside your existing application. Give it the current documentation, a clear business task and access to the workspace you want it to use.
For example, ask it to build a service desk that reads delivery guidance from your backend. It can define the conversation, implement the tool, write tests and explain the proposed publication. You review the behavior before connecting a phone number.
Connect the documentation
Connect the Cumulus MCP for your client. The public documentation server needs no key:
https://mcp.cumuluslabs.io/mcp/docs
The assistant can search guides and read exact request and response schemas. Ask it to read the platform overview, then the TypeScript tutorial. If it cannot use MCP, give it the Markdown index.
Give it a useful first task
Start with the outcome you want and the system you already use:
Build a service-desk voice agent in this application. It should ask whether the
caller needs delivery or repair information, call our backend for guidance,
explain the result and end politely. If the lookup fails, say so.
Read the current Cumulus TypeScript tutorial and use the installed SDK 0.3.4.
Keep the tool handler in our server. Use fictional data for the first version.
Run the local tests and CLI validation, then explain the files you created.
Do not publish, assign a number or call anyone yet.
Local work needs the SDK package and Node.js. Publishing also needs your development workspace's API URL and key. Your backend must be reachable over HTTPS before a remote agent can call its tool.
Add workspace access
The authenticated MCP connection lets the assistant read and change resources allowed by its API key. Client setup includes Codex and Claude Code commands. For diagnosis, use a viewer key; use a member key when you want it to edit agents or start runs.
Before asking for changes, verify access:
Identify my current environment and list my agents without changing anything.
Then compare the local service-desk definition with its remote draft and show
me the proposed changes.
Review the workspace and preview. Applying edits, publishing a version, assigning a number and dialing are separate steps. Release an agent explains the full sequence.
Add project instructions
Put the conventions you want repeated in your application's AGENTS.md or CLAUDE.md. Adapt this example:
Use the current Cumulus documentation and installed @cumulus/sdk package.
Read the relevant guide and exact operation schema before changing integration code.
Keep agent definitions in source control and tool implementations in our backend.
Use stable node/tool keys, an explicit failure path and durable tool-effect receipts.
Run local tests and CLI validation before previewing remote changes.
Confirm the development workspace and review the preview before applying a draft.
Keep credentials and personal numbers out of source, prompts and recordings.
Use dedicated test numbers; test phone runs still make real calls.
These instructions help the assistant follow your development process. Your tool handlers still enforce access to business records and operations; see customer HTTP adapters.
Review what it built
Check that the conversation accomplishes the task, the backend returns real structured results and failures lead somewhere useful. A definition passing validation is the starting point. Test a successful lookup, a failed lookup and a repeated tool request, then inspect the result of a dedicated test run.
The complete example includes those local checks. Investigate a run shows how to review the conversation, tools and outcome.
Send a page to your assistant
Each page offers Open in Codex, Open in Claude, Copy agent prompt and View Markdown. These carry the page and section into your conversation. Add what you want to build or understand so the assistant has a concrete task.
Troubleshooting
If it invents an SDK method, have it inspect the installed package and read the matching reference. If documentation works but workspace access fails, check the second MCP connection and its key. MCP troubleshooting covers connection errors; SDK errors covers request recovery.