v0.1.0

oc3

A minimal local proxy that lets OpenAI Codex and ChatGPT desktop use OpenCode Console models — device-code OAuth, Responses-API bridging, hosted web search, and native OpenAI fallback through one local endpoint.

Install

npm install -g @nbrst/oc3

curl -fsSL https://oc3.nbr.st/install | bash

Standalone Bun binaries for macOS (arm64/x64) and Linux (x64/arm64) — no runtime required. Or clone and bun install.

Run

bun src/cli.ts start    # overrides config, boots proxy + ChatGPT desktop
bun src/cli.ts stop     # restores previous config, stops proxy

How it routes

responses — GPT, Grok, Muse models on Console

# Near-passthrough: OAuth + x-opencode-* headers injected,
# SSE relayed verbatim between Codex and Console

chat-completions — everything else

# Full translation: Responses requests -> Chat Completions,
# streaming events re-emitted Codex-side (tool calls, reasoning,
# truncated stops, usage details)

messages / google — Claude and Gemini on Console

# Bridged via Anthropic Messages and Gemini GenerateContent
# translators (thinking, tool_use, functionCall, usage)

web_search — hosted tool on any model

# Codex's hosted web_search tool is bridged client-side via
# Exa/Parallel MCP, mirroring OpenCode's own implementation

Features

Device-code OAuth with single-flight token refresh and org switching
Schema-aware tool-argument repair (integral floats, string-typed ints)
Strict usage details on every response (strict Responses clients)
Routing catalog with per-model thinking/effort normalization
Auto-review alias resolution, Full-Access exec schema scrubbing
Native OpenAI fallback with client auth (ChatGPT backend or API key)
ChatGPT desktop transport: 426 HTTP fallback, zstd request bodies
OpenTUI dashboard: model picker, server + config override toggles
Detached daemon with oc3 status / oc3 logs / one-command restore

How it routes

Features

Links