Your phone agent now has regression tests, and a robot caller to run them

What it does

Lampline is a self-hosted platform for building and running phone agents. Agents are plain folders of YAML and Markdown — reviewed in git, linted and tested by the vk CLI — and the same agents take real calls over Twilio, Telnyx, or any SIP trunk via jambonz. The twist is that it treats conversation behavior as testable code: you write YAML tests specifying what the caller says, which steps the call must reach, what the agent must (or must never) say, and which tools it calls.

The interesting bit

The Evals tab actually dials your agent on a real phone line and plays the caller with an AI Test Caller — background noise, accents, interruptions, rambling included — then grades the transcript check by check. Free calls ride through a Google Voice browser extension (you press Call yourself; Google Voice forbids automated dialing). It even auto-rewrites failing prompts, keeping a change only if more checks pass and nothing that passed breaks. There’s also a two-way Dialogflow CX / CX Agent Studio bridge: import an export, edit locally, export back.

Key highlights

• Agents as code: agent.yaml plus prompts/*.md, tested with vk test from terminal or CI
• Skills for Claude Code, Cursor, and Codex to build, fix, and review agents; vk mcp exposes agent folders as tools
• Runs fully local or on-prem, no telemetry, bound to localhost, keys kept out of agents and the repo
• 20 model providers, or Ollama on your own hardware; free to run with no API keys
• Imports Hammer, Cyara, or spreadsheet test scripts as runnable cases; replays failed calls
• Measures response time as callers hear it, phone line included (typical and p95)

Caveats

• The Dialogflow import is documented as incomplete — the README points to a list of what doesn’t carry over yet, and webhooks answer with mocks until connected
• Real-call evals depend on the Google Voice bridge or a phone on speaker, which is a clever workaround but not exactly an enterprise dialing rig
• Built on Pipecat, so you inherit that stack’s opinions along with it

Verdict

If you run contact-center or IVR workloads — especially anything touching Dialogflow CX — this is a rare combination of test discipline and self-hosting. If you just want a hosted voice bot in ten minutes, it’s more machinery than you need.