Lampline is an open-source platform for self-hosting Dialogflow CX agents

A voice agent built in Google's console was something a team could not take with it. Lampline, now at 145 GitHub stars in its first week, imports a Dialogflow CX or CX Agent Studio export as a drop-in zip and runs it on your own machine as a drag-and-drop builder, a test console and a phone endpoint.

The project, by Rohan Salvi, is published under Apache-2.0 and built on Pipecat, a BSD-licensed voice pipeline. The import, builder and phone-call pieces are already marked done on the roadmap.

What the import carries over

Lampline's own README says a Dialogflow CX export brings over flows, pages, form parameters and entity types, intent and condition routes, no-match handlers, playbooks, webhooks and OpenAPI tools. CX Agent Studio exports bring over the root agent and child agents, global instructions, variables, OpenAPI toolsets, Python tools and data stores. Pages become scripted pages said word for word, playbooks become LLM steps, webhooks become tools and the data stores become local knowledge search. Export the other way is also claimed, writing any Lampline agent back as a Dialogflow CX package or a CX Agent Studio app.

Twenty model providers, and local models for speech

The Integrations page lists Claude, OpenAI, Gemini, Vertex AI, Azure OpenAI, Amazon Bedrock, Groq, Mistral, DeepSeek, Grok, Together AI, Fireworks AI, OpenRouter, Cerebras, NVIDIA NIM, Qwen, SambaNova, Nebius, Novita AI and Ollama. The project also bundles Whisper, Moonshine, Parakeet, Kokoro and Qwen3-TTS so speech can run free on the operator's own hardware. Calls arrive over Twilio, Telnyx or any SIP trunk through jambonz, with the webhook to copy on the Phone numbers page.

The pricing comparison is the project's own, at list prices

Lampline publishes a cost table with CX Agent Studio's list price as the baseline. It puts 10,000 four-minute calls a month at $5,000 on CX Agent Studio against $1,384 on an on-prem server using Google model APIs, and 200,000 calls at $100,000 against $24,426 on the same stack. Self-hosted models on the same server fall to $681 at 10,000 calls and $3,748 at 200,000. These are Lampline's own figures at list prices for October 2026 and the project says staff time is not included, two assumptions nobody has published good data for (concurrent calls per GPU, and whether an 8B open model is good enough for a given flow) are unmeasured, and the model APIs are almost the whole bill. The repository ships python3 costs/cost_model.py to reproduce or change any assumption.

Agents as code, with limits the project states openly

Every agent is a folder of YAML and one Markdown file per prompt, so changes are reviewable in git. A vk CLI runs lint, conversation tests, templates and an installer for Claude Code skills. The engine reads the folders on every call, so a change is live on the next call. The README names what the project does not yet do: a Google Cloud deployment and connectors for Genesys Cloud, Avaya, Cisco Webex Contact Center, Cisco UCCE, Amazon Connect, NICE CXone, Five9 and Twilio Flex are marked coming soon, published benchmarks are still open, and the project says it is unaffiliated with Google or any provider it names. The engine starts with one Docker Compose command on a laptop, a GPU box or an on-prem server.

Source

Lampline on GitHub, by Rohan Salvi.