Documentation
Documentation
A private club for local inference. Share models with a trusted group over a libp2p mesh, behind one OpenAI-compatible endpoint.
Speakeasy is a Go inference proxy for people who already run local models (Ollama today) and want a trusted group to use them without exposing those backends to the internet.
Each node keeps its own weights. Members join a private p2p mesh. A request hits the local proxy first; if that model is not here, the proxy forwards to a peer that advertised it. All inference traffic is peer-to-peer and encrypted (libp2p QUIC). The relay is only for setup and hole punching — it never carries chat.
Install
Prerequisites, build, and the files the binary writes.
Quick start
Join with an invite, or stand up your own mesh.
Admin setup
Stand up a seed/admin node, put TLS in front of the admin API, and issue invites.
How it works
Admin, relay, local-first routing, and what never leaves the node.
Configuration
Runtime settings live in config.yaml in the process working directory.
Usage
Point Ollama CLI, OpenAI SDKs, and Open WebUI at the local proxy.
Dashboard
Welcome, Mesh, Nodes, Models, Chat, Admin, and Settings. Default listen serves the UI at /.
CLI
speakeasy for join, proxy, and admin. admincli for invites and nodes.
API
Admin HTTP API, dashboard JSON, and peer-only mesh RPC.