Portage

Portage

**Keep the harness. Change the model.**

[](https://github.com/SimonGino/portage/actions/workflows/ci.yml) [](https://go.dev) [](#quick-start)

[English](README.md) · [简体中文](README.zh-CN.md)

Portage is a self-hosted model gateway: put every model you can reach — official APIs, OpenAI-compatible relays, Ollama / vLLM on your own machine — behind one address, and run any of them inside agent harnesses like Claude Code and Codex CLI.

It works by translating protocols. Each harness speaks exactly one (Claude Code speaks Anthropic Messages, Codex CLI speaks OpenAI Responses, while most models only expose OpenAI Chat Completions). Portage translates between all three in the middle, and passes bytes through untouched when the protocols already match. Point the harness at Portage and switching models is one line of config — no patched clients, no forked harness, no wrapper scripts.

flowchart LR
    subgraph clients["The harness you already use"]
        CC["Claude Code
Anthropic Messages"]
        CX["Codex CLI
OpenAI Responses"]
        APP["Your scripts / SDK
Chat Completions"]
    end

    PG{{"Portage
one binary · SQLite · optional web admin"}}

    subgraph up["Whatever model you can get"]
        T["Open-weight models
via any OpenAI-compatible relay"]
        L["Your own hardware
Ollama · vLLM · MLX"]
        A["Anthropic · OpenAI
native"]
    end

    CC --> PG
    CX --> PG
    APP --> PG
    PG -- "different protocol → translate" --> T
    PG --> L
    PG -- "same protocol → byte passthrough" --> A

What you can run

The model you want	Reached over	In the harness
Open-weight models — whatever the good one is this month	any OpenAI-compatible end