C4rohan/Lampline

Bring your agents over from Dialogflow CX and CX Agent Studio, or start fresh.
Build, test and run them on your own machine or servers, in one command,
for about $0.02–0.17 a call instead of $0.50.

🚀 Quick start  ·  📋 Free phone evals  ·  🔁 Import from Google  ·  👀 Tour  ·  🧠 Models  ·  📞 Phone and SIP  ·  🗺️ Roadmap  ·  📘 Engine docs  ·  🧩 vk CLI

_{Import a Dialogflow CX agent, test it, and grade a real phone call to it. Sample agent and made-up data; replies from Lampline's test model.}

Lampline imports your Dialogflow CX and CX Agent Studio agents, flows, pages, playbooks, forms and tools included, and runs them on hardware you control. It gives you a drag-and-drop builder where LLM steps and scripted pages mix in one agent, a test console that lights up each step as the call moves, HTTP and MCP tools, records and document search of its own, and phone calls over Twilio, Telnyx or any SIP trunk through jambonz. Built on Pipecat.

Lampline's Evals tab phones a voice agent over a real phone line, plays the caller with an AI Test Caller, and grades the call against what should happen. It works on any agent with a phone number: one built in Lampline, Dialogflow CX, CX Agent Studio, Vapi, Retell, or your IVR.

Free calls through Google Voice. The Lampline Phone Bridge extension (Chrome or Edge) connects the call's audio to the Test Caller inside the browser: no cables, Windows or Mac. Or use any phone on speaker. Graded, with reasons. You write who is calling, what they want and what must happen; your model grades the transcript, check by check. A call that heard nothing shows Didn't run, never Failed. Tough callers. Background noise, accents, interruptions, long pauses, hesitant or rambling speech. The agent's speed, as callers hear it. Typical and p95 response time, phone line included. Bring your test scripts. Paste Hammer, Cyara or spreadsheet tests and they become runnable cases. Replay any failed call. How to run evals →

Only call agents you own or are allowed to test. You press Call yourself: Google Voice doesn't allow automated dialing.

Export an agent from Google's console, drop the .zip on New agent → Import, and it opens in the builder: flows, pages, forms, routes, playbooks and tools carried over, with a report of anything to check before it takes calls. It works the other way too: Export writes any agent back as a Dialogflow CX package or a CX Agent Studio app.

📥 Import. A Dialogflow CX package or a CX Agent Studio app (JSON or YAML). The conversion runs on your computer and never runs code from the export.

✏️ Ready to edit. Pages become scripted pages, playbooks become LLM steps, flows group them, and webhooks and OpenAPI tools become tools.

From the terminal: vk import auto export.zip and vk export dfcx|ces agents/<name>. The full mapping, both ways, and what doesn't carry over yet: cli/README.md.

_{Sample data only.}

🧪 Test calls. Chat or talk through your microphone. The steps light up as the call moves, and every collected detail shows on the right.	🧠 20 model providers. Pick one, paste a key, press Test connection. Speech runs in the cloud or free on your own computer.
📞 Phone calls and SIP. Twilio, Telnyx, or your own SIP trunk, PBX or SBC through jambonz. Genesys, Avaya, Cisco and more are coming.	🗄️ Data of its own. Collections agents look up and save during calls, and document search over your FAQs, all on your machine.

_{Screenshots use the bundled sample agents and made-up data.}

Needs Docker with Compose v2 (Docker Desktop is fine). No API keys or accounts:

git clone https://github.com/C4rohan/lampline && cd lampline
scripts/docker-up.sh --free          # everything on this computer: Ollama, Moonshine and Kokoro

--free runs the model (Qwen3 4B in Ollama) and speech (Moonshine listening, Kokoro speaking) on your own computer. The first start downloads about 3 GB and shows progress; after that it starts in seconds. Replies are slower and simpler than a cloud model's, and you can switch to any provider on the Integrations page whenever you like. On a Mac, the native route is faster, because Ollama uses the GPU outside Docker: install Ollama, then scripts/engine-up.sh --free.

Have API keys? Run scripts/docker-up.sh without --free. Either way it prints a link like http://localhost:7860/#/?key=…: open it once and the page remembers the key. Pick your model and voice on the Integrations page (or put keys like ANTHROPIC_API_KEY, GEMINI_API_KEY, DEEPGRAM_API_KEY in .env; see .env.example). Your agents are in ./agents (edit them, keep them in git) and everything else the engine keeps, settings with your keys, records, knowledge and downloaded models, is in ./.vpe-data. scripts/docker-up.sh --down stops it; --logs follows its log. Plain docker compose up -d --build works too once .env has VPE_DOCKER_API_KEY. Details: engine/README.md.

Prefer no Docker? scripts/engine-up.sh runs the engine natively (Python 3.11+) at http://localhost:7860, which also lets browser calls reach it directly. Free, private speech: Integrations → Whisper and Kokoro → Download models, or Moonshine for listening, which transcribes while the caller talks.

Pick one on the Integrations page; each starts on a sensible default model you can change, and Test connection checks the key.

The Phone numbers page has the webhook to copy and the steps for each. Details: engine/README.md.

Genesys Cloud, Avaya, Cisco Webex Contact Center, Cisco UCCE and PCCE, Amazon Connect, NICE CXone, Five9 and Twilio Flex show on the Integrations page as coming soon. Until each lands, most of them can already send calls to your agents over a SIP trunk through jambonz. The plan for each: docs/ROADMAP.md.

flowchart LR
  you((You)) --> app[Lampline app<br/>builder, tests, integrations :7860]
  app --> engine[Engine<br/>Pipecat pipelines]
  phone((Callers)) --> carriers[Twilio · Telnyx ·<br/>SIP via jambonz]
  carriers --> engine
  engine --> agents[(agents/<br/>YAML + Markdown)]
  engine --> data[(Records and<br/>knowledge, SQLite)]
  engine --> models{{Model and speech<br/>cloud APIs or local<br/>Whisper · Moonshine · Parakeet · Kokoro · Qwen3-TTS · Ollama}}
  engine -- tools --> hooks[Your APIs and<br/>MCP servers]

The engine reads the agent folders on every call, so a change is live on the next call. Settings and keys live in a chmod 600 file outside the repo.

Monthly totals for 4-minute calls at list prices (October 2026). Staff time is not included. Reproduce or change any assumption with python3 costs/cost_model.py.

What the numbers say:

• 📉 Hosting is the small part. Model APIs are almost the whole bill, so the model choice matters far more than local vs. cloud.
• 🖥️ Self-hosted models are where the big savings are, but they rest on two assumptions nobody has published good data for: concurrent calls per GPU, and whether an 8B open model is good enough for your flows. Help measure them.

Every row, per-call figures and assumptions: costs/COSTS.md.

Build and change agents from a terminal or an AI coding agent as well as the builder. Each agent is a folder: agent.yaml for the graph and one Markdown file per prompt, so changes are reviewable in git.

pip install -e cli                  # installs the `vk` command
vk templates install front-desk     # start from a ready-made agent
vk lint agents/front-desk           # graph checks + a scan for real-looking patient data
vk test agents/front-desk           # runs tests/*.yaml conversations on the engine
vk import dfcx export.zip           # bring an agent over from Dialogflow CX or CX Agent Studio
vk skills install --target claude   # builder + reviewer skills for Claude Code

Details in cli/README.md. The engine itself, its builder, tools, data and knowledge: engine/README.md.

• The engine on Pipecat: builder, text chat, browser calls, phone calls (engine/README.md)
• Agents as code: vk lint, conversation tests, templates, coding-agent skills (cli/README.md)
• Scripted pages and flows mixed with LLM steps (engine/README.md)
• Dialogflow CX and CX Agent Studio import and export (cli/README.md)
• Records and local RAG of its own (engine/README.md)
• One Docker container for local and on-prem
• 20 model providers; SIP through jambonz
• Transfer to a person with the call's details as SIP headers (engine/README.md)
• Google Cloud deployment (coming soon)
• Contact centers: Genesys Cloud, Avaya, Cisco Webex Contact Center, Cisco UCCE, Amazon Connect, NICE CXone, Five9, Twilio Flex (coming soon)
• Users and roles, and an HTTPS guide (docs/HTTPS.md)
• Test a phone agent for free from your laptop (docs/PHONE-TESTING.md)
• Evals: real phone calls through Google Voice, played by a Test Caller and graded (docs/EVALS.md)
• Import Hammer, Cyara or spreadsheet test scripts as eval test cases (docs/EVALS.md)
• Evals placed by Lampline through Twilio, for unattended runs
• Call history and optional recordings, on this computer (engine/README.md)
• Published benchmarks

The full list: docs/ROADMAP.md.

CI runs on every push and pull request to dev, qa and prod:

• the vk CLI test suite
• the engine test suite: flows, extraction, text chat, evals, every model provider's setup, Twilio, Telnyx and jambonz calls, the voice pipeline and a real WebRTC handshake, on fake models
• the web app builds, and the committed build matches its source
• the Docker image builds, starts as a non-root user, answers health and refuses requests without its key
• shellcheck, and a check that COSTS.md matches the cost model
• a scan of every commit for keys and tokens (gitleaks)

Issues and PRs welcome (against dev; how dev, qa and prod work). See CONTRIBUTING.md, SECURITY.md, the code of conduct and the changelog.

Lampline is created and maintained by Rohan Salvi. If you use it in a project, product, paper or talk, please credit "Lampline by Rohan Salvi" and link this repository. GitHub's Cite this repository button (from CITATION.cff) gives APA and BibTeX:

@software{salvi_lampline_2026,
  author  = {Salvi, Rohan},
  title   = {Lampline: open-source, self-hosted voice agents},
  year    = {2026},
  version = {0.3.0},
  url     = {https://github.com/C4rohan/lampline}
}

If you redistribute Lampline or build on it, the Apache-2.0 licence requires you to keep the NOTICE file, which carries this attribution.

Apache-2.0. The engine uses Pipecat (BSD-2-Clause), installed from PyPI; other credits are in NOTICE. Not affiliated with Google or any provider named here. Prices are list-price estimates that change often; check them before budgeting.