We open-source the infrastructure we build for ourselves and our clients. Everything below runs in production at our home lab — take it, run it, break it, improve it.
Stop juggling connection settings in every tool. El Pulpo sits between your clients and your inference boxes: one stable OpenAI-compatible endpoint, one model catalogue that doesn’t change when your fleet topology does, and a web dashboard that tracks token usage and the money you save against cloud prices. Fleet changes are a single edit instead of a hunt through every agent and harness.
One stable endpoint & model catalogue for all your LLM servers
Centralized server & pricing configuration, applied without restarts
Usage accounting and savings dashboard per request
Single static binary, 12.7 MB multi-arch container image
All your inference servers, configured in one place.
An audio service that speaks the OpenAI /audio API: speech-to-text, text-to-speech, and model listing — backed by ONNX Runtime so it runs on CPU or GPU with no PyTorch at runtime. Anything already built for OpenAI’s audio endpoints works against it with a base-URL swap.
Drop-in /v1/audio/speech and /v1/audio/transcriptions
Whisper ASR in ~99 languages; Kokoro & Piper voices for TTS
Pull models from the Hugging Face registry and hot-load them
Docker images for CPU and CUDA; runs happily on small hardware
Talk instead of typing, and listen instead of reading. A browser-side client plugin for the DeepSeek Harness web GUI: a microphone button in the composer dictates your draft via whisper, and a speaker button on every assistant message reads it back via Kokoro. Wired to Camaron Audio by default, and it works against any OpenAI-compatible audio server.
Dictate: record in-browser, transcribe, review, then send
Listen: one-voice-at-a-time playback on every assistant reply
Hand-authored single file — the source is the artifact, no build step
Uses only public plugin slots and hooks; no internals touched
More coming out of the tank
Follow the organization on GitHub for new releases, or reach out if you want a project adapted to your stack.
