Set up and manage your voice agent from a coding agent over MCP.

How do you create a token for your coding agent?

Sign in, open Developers and choose Create access token. The token is shown only once, so add it to your MCP client straight away, ideally through an environment variable rather than a file you commit.

A token has full management access to one workspace. It can change your agents and knowledge, approve sources, read your conversations and create or revoke other tokens, so create one for each tool you connect and revoke any you stop using. Kawl Now emails your sign-in address when a token is created or revoked.

How do you connect your MCP client to the server?

Your MCP client needs to support remote servers over Streamable HTTP and let you add an Authorization header. Give it the three settings in the first example below, and check its documentation for where they go.

The server supports protocol versions 2026-07-28 and 2025-11-25. Kawl Now’s automated tests cover both, but they don’t run inside any named client. The Claude Code and Cursor examples follow each client’s own documentation, reviewed on 8 October 2026, so check the connection before you rely on it.

In Claude Code, run the command from your project folder with your token in place of YOUR_WORKSPACE_TOKEN. Claude Code adds servers at local scope by default, which keeps the entry in ~/.claude.json for that project only, so the token stays out of your repository’s shared .mcp.json. Run claude mcp get kawl-now, or /mcp inside Claude Code, to see whether it connected.

In Cursor, add the entry to ~/.cursor/mcp.json and set KAWLNOW_TOKEN as an environment variable that Cursor can read. Cursor resolves the variable when it loads the file, so the file itself holds no secret.

The Developers tab also generates a JSON config with your new token filled in, so keep that copy out of files you commit. Claude Code needs "type": "http" added to an entry like that. Whichever client you use, ask your coding agent to call get_workspace first: if it returns your agents’ settings, the connection works.

Server URL:  https://app.kawlnow.com/api/mcp
Transport:   Streamable HTTP
Header:      Authorization: Bearer YOUR_WORKSPACE_TOKEN

claude mcp add --transport http kawl-now https://app.kawlnow.com/api/mcp \
  --header "Authorization: Bearer YOUR_WORKSPACE_TOKEN"

{
  "mcpServers": {
    "kawl-now": {
      "url": "https://app.kawlnow.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ${env:KAWLNOW_TOKEN}"
      }
    }
  }
}

How does your coding agent add knowledge for review?

add_knowledge saves a support or sales document of up to 50,000 characters, and sync_github imports one .md, .mdx or .txt file from a public GitHub repository into support knowledge. Both save the source unapproved, so it shows as Needs review in the dashboard, and a workspace holds up to 100 sources.

Your coding agent can then call list_knowledge to see what’s waiting and get_knowledge to read a source’s full text back to you. Each GitHub re-sync makes that source unapproved again, even when nothing changed. Website import is switched off and the import_website tool returns an error, so add pricing and product page copy as sales notes with add_knowledge.

Ask in plain language and say which steps you want to keep for yourself. For example:

• “Add the answers in docs/faq.md as support knowledge, then list what still needs approval.”
• “Sync docs/setup.md from example-org/product on GitHub and show me its full text.”
• “Turn our pricing page copy into a sales note and leave it unapproved for me to check.”

How do you keep approval with a person?

Adding a source never approves it, whether it arrives through MCP, the REST API or the dashboard. Your voice agent searches only approved sources, so no knowledge your coding agent adds reaches callers until someone approves it. Settings it changes with update_agent, such as the greeting and instructions, apply straight away, so review those changes yourself too.

Any workspace token can still approve sources, through approve_knowledge or the REST API. To keep that decision with a person, approve sources yourself in the dashboard and stop your coding agent calling approve_knowledge. If your MCP client lets you block a tool or confirm each call, use that.

In Claude Code, the deny rule shown here removes the tool, so Claude can’t call it. Add it with /permissions or in a settings file such as .claude/settings.json. The rule names the server as you added it: this guide and the Developers-tab config both call it kawl-now, so rename the rule if you chose another name. Cursor asks before it runs an MCP tool by default and lets you check each call’s arguments, so you can decline approve_knowledge calls, although its run modes can change that default.

Blocking the tool covers MCP only. A coding agent that can run commands and read your token, for example from an environment variable, could still approve sources over the REST API, so check which sources are approved in the dashboard yourself.

{
  "permissions": {
    "deny": [
      "mcp__kawl-now__approve_knowledge"
    ]
  }
}

How does your coding agent set up the agent and the widget?

update_agent changes the product and agent names, greeting, instructions, voice, button label, colour, allowed website origins and whether customer calls are on. It can also switch human handover, an early release, on or off once you’ve verified that agent’s team phone number in the dashboard. It works on the support agent unless you add channel: "sales". Changes apply straight away, including switching customer calls on, so ask your coding agent to leave calls off until you’ve tested each agent in the dashboard.

Allowed origins are saved as a whole list, so the origins update_agent sends replace the saved ones. Ask your coding agent to read the current list with get_installation and send it back with any new origin, because any origin it leaves out stops working.

When a token switches calls or handover on or off, or changes the allowed origins, Kawl Now emails your sign-in address and names the token by its first characters.

get_installation returns the support or sales embed code with its allowed origins and whether calls are on, so your coding agent can paste the snippet just before the closing </body> tag of the pages that need it. You might ask:

• “Add https://www.example.com to the sales agent’s allowed origins, keeping the existing ones, then put the sales embed on our pricing page.”
• “Set the support agent’s greeting and instructions from docs/voice.md, and leave customer calls off.”

How do you check conversations from your coding agent?

list_conversations returns up to 100 recent conversations with each caller’s topic, call status, length, whether it was a dashboard test and any feedback, plus the start of what the caller said once the call has ended.

conversation_insights returns the automatic insights for the support agent, or the sales agent with channel: "sales". They group recurring questions, reported issues and themes from what callers said in up to the latest 30 completed customer calls in the last 30 days, with exact quotes, and dashboard tests aren’t included. Ask your coding agent to suggest which docs to add or fix, then review the new sources like any others.

Full transcripts aren’t available over MCP. Read them in the dashboard or fetch them with the REST API.

What do you still do in the dashboard?

Create your first token and run test calls in the dashboard, because neither is available over MCP. Test each agent you plan to use before you switch on customer calls. The first 50 conversations in a workspace are free and shared by support, sales and tests. Paid calls aren’t available yet, so new calls stop once they’re used up.

Common questions