Projects / ailocal

ailocal

activev0.9.1

Runs Claude Code, Codex CLI and VS Code Copilot Chat against models on your own Mac. Detects the clients you already have and configures all of them from one source of truth.

2releases

ailocal exists because the two costs of cloud coding assistants — per-token billing and source code leaving the machine — are both avoidable, and neither is worth paying if the tooling does not have to change.

It configures three things for you: the models, the local API they are served on, and every supported client that is present. It deliberately does not install software. You install the prerequisites and whichever clients you want; ailocal detects what is there and wires it up. That boundary is the whole design — a tool that installs things owns your machine, and a tool that configures things can be removed cleanly.

Architecture

ailocal request pathThree unmodified clients — Claude Code, Codex CLI and VS Code Copilot Chat — speak either the OpenAI or the Anthropic dialect to a LiteLLM proxy on 127.0.0.1:4000. The proxy exposes capability names rather than model tags. A hardware profile resolves each capability to a model, and LiteLLM forwards the request to Ollama running on the host GPU.CLIENTS — UNMODIFIEDClaude CodeCodex CLIVS Code CopilotAnthropicOpenAIOpenAICOMPATIBILITY SEAMLiteLLM proxy127.0.0.1:4000exposes capabilities,not model tags:architecture · review · fastimplementation · embeddingsRESOLUTIONprofiles/<tier>.tomlcapability → model, geometryresolved modelBACKENDOllamaon the host,not a containerhost GPUOne seam. Clients keep their own dialect and are never modified.
Fig. 1 — The request path. Clients keep their own API dialect; the proxy is the only component that has to know both, and the hardware profile is where a capability name becomes a model.
  1. Aug 08, 2026releaseContext windows sized from measured runner behaviourv0.9.1
  2. Aug 07, 2026releaseFirst public releasev0.9.0
ESC
↑↓ navigate↵ open