Claude Proxy

Run Claude Code on any model — GPT-5.6 Codex, Gemini 3, GLM-5.2, Claude, or anything on OpenRouter — without changing how you work.

Claude Proxy is a tiny local server that speaks the Anthropic Messages API. Claude Code keeps talking to http://127.0.0.1:17870/v1/messages; the proxy translates each request into the target provider's native format, streams the reply back as Anthropic SSE, and lets you hop between models mid-session with /model.

Log in with the accounts you already pay for — a ChatGPT/Codex plan, a Google account, a Cline subscription — or bring plain API keys. No Anthropic key required.

Codex (OAuth) — GPT-5.6 Sol / Terra / Luna and GPT-5.5, with per-model reasoning effort
Gemini (OAuth) — Gemini 3.1 Pro / Flash, 3 Pro, and 2.5, via Google Code Assist
ClinePass — every model in your Cline subscription at $0 per call
API keys — OpenAI Responses, OpenRouter, Google Gemini, Z.AI GLM, Anthropic passthrough
Streaming + non-streaming, tool use, vision, thinking blocks, and native server-side web search

There's no package to install from a registry — you clone the repo and run it locally. Point your coding agent at it and it can set everything up for you; see CLAUDE.md.

Setup

Prerequisites: Node 18+, git, and Claude Code already installed.

git clone https://github.com/aryan877/claude-proxy.git
cd claude-proxy
npm install     # runtime deps (fastify, tsx, …)
npm link        # exposes the launcher commands on your PATH

Then start a launcher — it boots the proxy, opens Claude Code wired to it, and stops the proxy when you exit:

claude-codex      # ChatGPT/Codex login  → GPT-5.6 Sol (High) by default
claude-gemini     # Google login          → Gemini 3.1 Pro
claude-cline      # Cline subscription     → GLM-5.2 ($0/call)
ccx --setup && ccx # your own API keys     → GLM-5.2

First run prints a login URL if you aren't authenticated yet.

Prefer not to npm link? Run a launcher directly, or alias it:

node ~/claude-proxy/bin/claude-codex.js
# or, in your shell rc:
alias claude-codex="node ~/claude-proxy/bin/claude-codex.js"

Commands

Command	Default model	Auth
`claude-codex`	`codex-oauth:gpt-5.6-sol@high`	ChatGPT/Codex OAuth, or existing Codex CLI tokens
`claude-gemini`	`gemini-oauth:gemini-3.1-pro-preview`	Google OAuth
`claude-cline`	`cline-pass:glm-5.2`	Cline subscription (token read from the Cline app)
`ccx`	`glm:glm-5.2`	API keys in `~/.claude-proxy/.env`

Each has a -d sibling (claude-codex-d, claude-gemini-d, claude-cline-d, ccx-d) that adds --dangerously-skip-permissions. All launchers share --status, --proxy-status, --restart, --stop, and -d; claude-codex and claude-gemini also support --logout.

After pulling changes, restart the proxy. A launcher reuses an already-running proxy, so new models/config won't apply until you run <launcher> --restart (or --stop, then relaunch).

Launchers

`claude-codex` — GPT-5.6 Codex via ChatGPT/OAuth

Routes Claude Code to OpenAI's ChatGPT/Codex backend. The GPT-5.6 family shares a 372k-token context window; Sol starts at High effort.

Shortcut	Model	Role
`sol` · `codex` · `cx`	`gpt-5.6-sol`	Frontier — hardest coding & research (default)
`terra`	`gpt-5.6-terra`	Balanced — everyday, high-volume work
`luna`	`gpt-5.6-luna`	Fast & affordable — routine tasks
`gpt55`	`gpt-5.5`	Previous frontier (272k window)

Tokens are looked up in ~/.claude-proxy/codex-oauth.json, then ~/.codex/auth.json. If neither exists, launch and open the login URL it prints:

http://127.0.0.1:17870/codex/login

`claude-gemini` — Gemini via Google OAuth

Routes Claude Code to Gemini through Google Code Assist. Default model is gemini-3.1-pro-preview, declared with a 1,000,000-token context window and Claude Code auto-compact disabled.

First run, open http://127.0.0.1:17870/google/login. Link a second Google account for automatic failover when the first hits a 429 (rate limit): http://127.0.0.1:17870/google/login/2. Tokens live in ~/.claude-proxy/google-oauth.json and …-oauth-2.json.

`claude-cline` — your Cline subscription ($0/call)

Routes Claude Code through your paid Cline / ClinePass subscription (https://api.cline.bot), so included models cost $0 per call. The OAuth token is read live from the Cline app (~/.cline/data/settings/providers.json) and refreshed by the proxy — keep the Cline app signed in. Default model is glm-5.2.

Shortcut	Model	Shortcut	Model
`cp` · `glm52`	`glm-5.2` (default)	`cpminimax`	`minimax-m3`
`cpkimi`	`kimi-k2.7-code`	`cpdeepseek`	`deepseek-v4-pro`
`cpkimi26`	`kimi-k2.6`	`cpflash`	`deepseek-v4-flash`
`cpqwen`	`qwen3.7-max`	`cpmimo`	`mimo-v2.5-pro`
`cpqwenplus`	`qwen3.7-plus`	`cpmimo25`	`mimo-v2.5`

Any Cline model also works verbatim: /model cline-pass:<id>. Set a session-wide thinking default with CLINE_REASONING_EFFORT (e.g. CLINE_REASONING_EFFORT=none claude-cline).

`ccx` — your own API keys, all providers

Starts the multi-provider API-key route. Claude Code launches on glm (glm:glm-5).

ccx --setup   # writes ~/.claude-proxy/.env
ccx

ccx --setup scaffolds ~/.claude-proxy/.env:

# OpenAI Responses API
OPENAI_API_KEY=
OPENAI_BASE_URL=https://api.openai.com/v1

# OpenRouter
OPENROUTER_API_KEY=
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
OPENROUTER_REFERER=
OPENROUTER_TITLE=Claude Code via ccx

# Gemini API key
GEMINI_API_KEY=
GEMINI_BASE_URL=https://generativelanguage.googleapis.com/v1beta

# Z.AI GLM
GLM_UPSTREAM_URL=https://api.z.ai/api/anthropic
ZAI_API_KEY=

# Anthropic passthrough
ANTHROPIC_UPSTREAM_URL=https://api.anthropic.com
ANTHROPIC_API_KEY=
ANTHROPIC_VERSION=2023-06-01

# Local proxy
CLAUDE_PROXY_PORT=17870

GLM_API_KEY is accepted as a fallback for ZAI_API_KEY.

Switching models

Use Claude Code's /model command. Shortcuts are defined in adapters/map.ts; explicit provider routes always work too.

/model sol            # a shortcut
/model codex-oauth:gpt-5.6-terra
/model gemini-oauth/gemini-3-flash-preview

Provider prefixes

Prefix	Route
`codex-oauth`	ChatGPT/Codex backend (OAuth)
`gemini-oauth`	Google Code Assist / Gemini (OAuth)
`cline-pass`	Cline subscription
`openai`	OpenAI Responses API (`OPENAI_API_KEY`)
`gemini`	Gemini API (`GEMINI_API_KEY`)
`openrouter`	OpenRouter (`OPENROUTER_API_KEY`)
`glm`	Z.AI Anthropic-compatible GLM
`anthropic`	Anthropic passthrough

Codex (GPT-5.6) shortcuts

Model shortcuts plus reasoning-effort shortcuts. Sol/Terra/Luna carry a 372k window; Sol and the other GPT-5.6 models default to High.

Shortcut	Route
`sol` · `codex` · `cx` · `gpt56` · `gpt-5.6-sol`	`gpt-5.6-sol@high`
`terra` · `gpt-5.6-terra`	`gpt-5.6-terra`
`luna` · `gpt-5.6-luna`	`gpt-5.6-luna`
`gpt55` · `gpt-5.5`	`gpt-5.5` (272k window)
`fast`	`gpt-5.6-sol@low`
`smart`	`gpt-5.6-sol@medium`
`deep`	`gpt-5.6-sol@high`
`xhigh` · `extra`	`gpt-5.6-sol@xhigh`
`max` · `think`	`gpt-5.6-sol@max`

All route through codex-oauth:. Append @level to override effort on any model — /model terra@max, /model luna@high, /model sol@low.

Gemini shortcuts

Shortcut	Route
`gemini` · `gemini-pro` · `gp` · `gemini-31p`	`gemini-oauth:gemini-3.1-pro-preview`
`gemini-flash` · `gf`	`gemini-oauth:gemini-3-flash-preview`
`gemini-3p`	`gemini-oauth:gemini-3-pro-preview`
`gemini-31f`	`gemini-oauth:gemini-3.1-flash-preview`
`gemini-25p` · `gemini-25f`	`gemini-oauth:gemini-2.5-pro` / `gemini-2.5-flash`

GLM, MiniMax, and Claude shortcuts

Shortcut	Route
`g` · `glm` · `glm52z`	`glm:glm-5.2` (current flagship, 1M ctx)
`glm51` · `glm5` · `glm47` · `glm45`	`glm:glm-5.1` / `glm-5` / `glm-4.7` / `glm-4.5`
`flash`	`glm:glm-4.7-flash`
`glm5or`	`openrouter:z-ai/glm-5`
`minimax` · `mm` · `m25`	`openrouter:minimax/minimax-m2.5`
`opus` · `sonnet` · `haiku`	`anthropic:claude-opus-4-8` / `claude-sonnet-5` / `claude-haiku-4-5`

Reasoning / thinking

Append an effort suffix to any supported model: @none · @minimal · @low · @medium · @high · @xhigh · @max.

Provider	How it maps
Codex / OpenAI	`reasoning.effort`. GPT-5.6 accepts up to `max`; GPT-5.5 tops out at `xhigh`. `none`/`minimal` floor to `low`.
Gemini 3.1 & Flash	`thinkingLevel`
Gemini 3 Pro	`LOW` or `HIGH` only (`medium` is raised to `HIGH`)
Gemini 2.5	thinking token budget
ClinePass	thinking level (`none`/`minimal` = off)

Defaults when you don't pass a level: Sol, Terra, and Luna → High; GPT-5.5 → High; Gemini → full thinking. Override the Codex fallback with CODEX_REASONING_EFFORT.

Why no ultra? In the Codex app, ultra is a multi-agent orchestration mode — it spawns subagents — and Codex itself sends it to the API as max. A single-request proxy has nothing extra to send, so max is the ceiling here.

Context windows

claude-codex declares a 372,000-token context window to Claude Code and auto-compacts at 95% by default. (GPT-5.5 is really 272k — set the override below for a 5.5-only session.) Claude Code's generated compaction-summary turn is automatically capped at Medium reasoning so a normal Extra High/Max session does not spend several minutes overthinking the summary; ordinary turns keep their selected effort.

CODEX_CONTEXT_WINDOW_TOKENS=272000 claude-codex   # e.g. when staying on gpt-5.5
CODEX_AUTO_COMPACT_WINDOW_TOKENS=340000 claude-codex
CODEX_DISABLE_COMPACT=1 claude-codex

claude-gemini declares a 1,000,000-token window and disables auto-compact by default.

How it works

Claude Code
    │  Anthropic Messages API
    ▼
Local proxy — 127.0.0.1:17870
    ├─ codex-oauth   → ChatGPT/Codex backend (OAuth)
    ├─ gemini-oauth  → Google Code Assist / Gemini
    ├─ cline-pass    → Cline subscription
    ├─ openai        → OpenAI Responses API (key)
    ├─ gemini        → Gemini API (key)
    ├─ openrouter    → OpenRouter
    ├─ glm           → Z.AI Anthropic-compatible GLM
    └─ anthropic     → Anthropic passthrough

The proxy handles:

Content blocks — text, images, tool use/results, thinking, redacted thinking
Streaming and non-streaming replies
Tool conversion for providers on OpenAI/Gemini tool schemas
Native server-side web search for Codex/OpenAI and Gemini (Claude Code's local WebSearch/WebFetch are stripped so the provider searches instead)
Reasoning cache — Codex encrypted-reasoning blobs are cached per session and replayed next turn so multi-step tasks keep their chain of thought
Internal-model remap — Claude Code's own haiku/sonnet probes (title generation, Explore subagent, quota checks) route to the active provider's equivalent, so nothing needs an Anthropic key
Vision — passthrough for Codex/OpenAI and Gemini; for GLM, images can be pre-described to text when OPENROUTER_API_KEY is set

Runtime files

State lives under ~/.claude-proxy/ (created on first run):

~/.claude-proxy/
├── .env                  # ccx API keys
├── codex-oauth.json      # ChatGPT/Codex tokens
├── google-oauth.json     # Google account 1
├── google-oauth-2.json   # Google account 2 (failover)
├── proxy.pid
└── proxy.log             # routing + upstream logs

Project layout

adapters/
├── anthropic-gateway.ts   # Fastify server + /v1/messages routing
├── map.ts                 # shortcuts + provider parsing
├── openai-auth.ts         # Codex OAuth
├── google-auth.ts         # Gemini OAuth
├── codex-reasoning-cache.ts
├── sse.ts · sse-aggregator.ts · vision-preprocess.ts · types.ts
└── providers/
    ├── codex-oauth.ts     # Anthropic → OpenAI Responses API
    ├── gemini-oauth.ts    ├── cline-pass.ts
    ├── openai-compat.ts   ├── openrouter.ts
    └── anthropic-pass.ts
bin/                       # launchers (claude-codex, claude-gemini, …) + lib/
tests/

Development:

npm test               # vitest
npm run start:proxy    # run the gateway directly (tsx)
npm run build          # tsc typecheck / emit

Troubleshooting

Proxy won't start / stale model. A launcher reuses a running proxy, so config changes need a restart:

claude-codex --restart          # or --stop, then relaunch
claude-codex --proxy-status     # shows the active provider:model
cat ~/.claude-proxy/proxy.log

If a non-proxy process holds the port, the launcher cleans it up via the PID lock and an lsof fallback.

OAuth looks stale. Tokens auto-refresh, but you can force a clean login:

claude-codex --logout && claude-codex
claude-gemini --logout && claude-gemini

ccx can't find keys. ccx --setup, fill in ~/.claude-proxy/.env, then ccx --restart.

/model didn't switch. Use a shortcut from this README or an explicit route (/model codex@high, /model openrouter:some-org/some-model), then watch the routing line: tail -f ~/.claude-proxy/proxy.log.

License

MIT — see LICENSE.

Name		Name	Last commit message	Last commit date
adapters		adapters
bin		bin
tests		tests
.gitignore		.gitignore
.npmignore		.npmignore
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
install		install
package.json		package.json
tsconfig.json		tsconfig.json

Navigation Menu

Repository navigation

About

Repository files navigation

Claude Proxy

Setup

Commands

Launchers

`claude-codex` — GPT-5.6 Codex via ChatGPT/OAuth

`claude-gemini` — Gemini via Google OAuth

`claude-cline` — your Cline subscription ($0/call)

`ccx` — your own API keys, all providers

Switching models

Provider prefixes

Codex (GPT-5.6) shortcuts

Gemini shortcuts

GLM, MiniMax, and Claude shortcuts

Reasoning / thinking

Context windows

How it works

Runtime files

Project layout

Troubleshooting

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Claude Proxy

Setup

Commands

Launchers

claude-codex — GPT-5.6 Codex via ChatGPT/OAuth

claude-gemini — Gemini via Google OAuth

claude-cline — your Cline subscription ($0/call)

ccx — your own API keys, all providers

Switching models

Provider prefixes

Codex (GPT-5.6) shortcuts

Gemini shortcuts

GLM, MiniMax, and Claude shortcuts

Reasoning / thinking

Context windows

How it works

Runtime files

Project layout

Troubleshooting

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

`claude-codex` — GPT-5.6 Codex via ChatGPT/OAuth

`claude-gemini` — Gemini via Google OAuth

`claude-cline` — your Cline subscription ($0/call)

`ccx` — your own API keys, all providers