Managing AI Voice Agents from the Terminal with Claude Code and MCP
Outcome: Complete voice-agent stack (agent config, LLM prompts, phone numbers, call review) managed from Claude Code sessions. Dashboard visits dropped to zero; iteration on agent prompts happens at conversation speed.
I wrote before about what missed calls cost a local business: 62% of small-business calls go unanswered, and most of those callers never call back. The fix is an AI voice agent that answers instantly. This is how I manage that stack without ever opening a dashboard.
The setup
Retell AI ships an MCP server. Plugged into Claude Code, every Retell operation becomes a tool the session can call: create_agent, update_retell_llm, create_phone_number, list_calls, get_call.
That turns voice-agent management into conversation:
Update the receptionist agent: if the caller asks about pricing,
collect their name and number, say Mark will call back within
the hour, and end politely. Then show me yesterday's calls
where the caller hung up in under 10 seconds.
Claude Code updates the agent’s LLM prompt through one tool call, then pulls and filters the call log through another. What used to be six dashboard screens is one instruction.
Why this beats the dashboard
Iteration speed. Voice agents live or die on prompt quality, and prompt quality comes from reading real transcripts. In a Claude Code session the loop is tight: pull transcripts, spot the failure pattern, rewrite the prompt, push the update, all in one place, with Claude doing the pattern-spotting across dozens of calls at once.
Version control. Agent prompts live as files in a git repo. The MCP server is just the deploy mechanism. Every prompt change is a commit; rolling back a bad change is git revert plus one tool call.
Compounding context. Claude Code sessions see the repo, the prompts, and the call data together. Ask “why do callers drop during the greeting?” and it reads the greeting prompt and the transcripts side by side.
The honest part
The voice agent itself is Retell’s product. I didn’t build the telephony. What Claude Code plus MCP replaces is the operational layer: the configuring, reviewing, and iterating that otherwise eats an hour of clicking per week.
That’s the pattern worth stealing, and it isn’t voice-specific. Any SaaS with an MCP server can be run this way: config as files in git, Claude Code as the operator, the dashboard as a thing you stop opening.