Bitdoze logo

VibeProxy Guide: Use Claude, Codex & Gemini in Any AI Tool

VibeProxy guide: use your Claude, Codex & Gemini subscriptions in any AI coding tool. Setup, supported models, Copilot, multi-account & troubleshooting.

Dragos

Updated Published 33 min read

VibeProxy Guide: Use Claude, Codex & Gemini in Any AI Tool

Disclosure: this guide contains referral links. If you buy through them, I may earn a commission at no extra cost to you.

You already pay for Claude Code, ChatGPT Plus, or another AI subscription. Why pay again for API access? VibeProxy is a macOS menu bar app that routes your existing subscriptions to coding tools like Factory AI Droids, Amp Code, Zed, Codex CLI, and any other OpenAI-compatible application. No API keys needed: OAuth in, model calls out.

Since the first version of this guide, the provider list has grown past Claude, Codex, and Gemini. VibeProxy now also connects GitHub Copilot, Kimi, and Z.AI GLM (Qwen was there before), and it can run several accounts per provider with automatic failover.

What you'll learn

  • Stop paying twice for AI by reusing the subscriptions you already fund
  • Install VibeProxy on macOS with verified DMG/ZIP downloads (Intel builds included)
  • Connect Claude Code, Codex, Gemini, GitHub Copilot, Kimi, Qwen, and Z.AI GLM
  • Configure Factory Droids, Amp Code, Zed, Codex CLI, and other tools against the local proxy
  • Run multi-account round-robin and pin providers with Priority toggles
  • Look up current model IDs for every supported provider (2026 reference tables)

What's new in this guide (updated Oct 2026)

VibeProxy ships near-daily (v1.8.341 as of 2026-10-21), and a lot changed since this article first went up:

  • Intel x86_64 builds exist since v1.8.58 (officially untested)
  • New providers joined: GitHub Copilot, Z.AI GLM (API key), Kimi (OAuth)
  • Multi-account round-robin with rate-limit failover, plus Provider Priority toggles
  • DMG + ZIP releases per architecture, each with a SHA-256 checksum file
  • Loopback-only binding enforced by default, with an opt-in LAN toggle
  • Vercel AI Gateway route for lower-risk Claude Max access
  • Refreshed model IDs across the board: Claude Opus 4.7, GPT-5.5 / 5.3-codex, Gemini 3.1

If you’re new to AI coding tools, start with our AI Programming Beginners Guide. If you already know the space and want the comparison table, see the best AI coding tools and agents in 2026.

What is VibeProxy?

VibeProxy is a free, MIT-licensed macOS app from Automaze (about 3,393 GitHub stars as of 2026-10-21). It runs in the menu bar and creates a local proxy on port 8317. Your coding tools talk to that proxy; the proxy talks to your existing subscriptions through the bundled CLIProxyAPI, which handles OAuth, token refresh, and routing. (The bundled binary is still named cli-proxy-api-plus, a leftover from the retired CLIProxyAPIPlus fork. Ignore the name.)

VibeProxy architecture: coding tools connect over loopback to the VibeProxy proxy on port 8317, which routes through the bundled CLIProxyAPI backend on port 8318 to subscription providers

Two local ports matter: 8317 is the proxy your tools point at, 8318 is the CLIProxyAPI backend behind it. Both bind to loopback. More on that in Security Considerations.

Why use VibeProxy?

  • Spend your existing flat-rate subscription quota instead of per-token API fees
  • One proxy serves many models: Claude Opus 4.7, Sonnet 4.6, GPT-5.5, Gemini 3.1, Copilot-hosted models, GLM-5
  • Browser OAuth per provider, so there are no API keys to manage; tokens stay local and refresh automatically
  • Several logins per provider give you round-robin distribution and failover when one account hits a rate limit
  • -thinking-N suffixes expose Claude reasoning budgets
  • Sparkle checks daily and pulls both VibeProxy and the bundled CLIProxyAPI forward
  • Anything that speaks the OpenAI or Anthropic API shape can point at it

[COST] VibeProxy itself is $0. The savings come from not double-paying: if you already fund a $20+/month Claude or ChatGPT plan, every API call through a third-party tool is money you do not spend twice.

Supported AI subscriptions: Claude, Codex, Gemini, Copilot, Kimi & GLM

VibeProxy talks to these providers today (source: project README and setup docs, fetched 2026-10-21):

Subscription Models available (current IDs) Auth Requirements
Claude Code Pro/Max claude-opus-4-7, claude-sonnet-4-6, thinking variants OAuth Active subscription
ChatGPT Plus/Pro (Codex) gpt-5.5, gpt-5.4, gpt-5.3-codex, gpt-5.3-codex-spark OAuth Active subscription
Google Antigravity gemini-3.1-pro-preview, gemini-3-pro-preview, Claude-via-Antigravity OAuth Google account
Gemini CLI gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite OAuth Google Cloud project
GitHub Copilot claude-opus-4.6, claude-sonnet-4.6, gpt-5.3-codex, gemini-3.1-pro-preview OAuth Active Copilot plan
Qwen qwen3-coder-plus, qwen3-coder-flash OAuth Qwen account
Kimi see in-app provider list OAuth Kimi account
Z.AI GLM glm-5, glm-4.7, glm-4-plus, glm-4-air, glm-4-flash API key GLM API key

Already have these subscriptions?

If you’re using Claude Sonnet and GPT models for free or hold any paid plan above, VibeProxy lets you spend that quota across every coding tool you own. GitHub Copilot subscribers get a second pool of Claude/GPT/Gemini usage through GitHub’s own API; see our GitHub Copilot Pro guide if you are still deciding on that plan.

[COST] Each provider draws its own quota. The Copilot route is a separate budget from your Claude Code or ChatGPT plan, which is exactly why it is useful: you can fail over between pools instead of burning one subscription dry.

The video above was recorded before the 2026 provider and UI additions (GitHub Copilot, Kimi, Z.AI GLM, Priority toggles). The setup flow is the same; the settings screen has more rows now.

Installing VibeProxy on macOS (DMG + SHA-256, Intel note)

System requirements

  • macOS 13.0 (Ventura) or later (LSMinimumSystemVersion in the app bundle)
  • Apple Silicon (M1/M2/M3/M4/M5) or Intel x86_64. Intel builds ship since v1.8.58 (2026-01-15) but are officially marked untested. Use them, report issues upstream.
  • An active subscription (or GLM API key) for at least one supported provider
  • Nothing else listening on ports 8317 (proxy) or 8318 (backend)

Prerequisites

  • macOS 13.0+ on Apple Silicon or Intel
  • At least one supported AI subscription
  • Ports 8317 and 8318 free on the machine

Quick pre-check for port conflicts:

bash
lsof -i :8317 -i :8318
# empty output = ports are free

Download and verify the release

Releases ship per architecture as both DMG and ZIP, each with a .sha256 file (latest DMG is roughly 34 MB as of v1.8.341, 2026-10-21):

File Use for
VibeProxy-arm64.dmg / .zip Apple Silicon Macs
VibeProxy-x86_64.dmg / .zip Intel Macs (untested builds)
*.sha256 checksum file next to each archive

Install with DMG

  1. Download VibeProxy-arm64.dmg (or VibeProxy-x86_64.dmg) from the releases page
  2. Verify the checksum:
bash
shasum -a 256 -c VibeProxy-arm64.dmg.sha256
# expect: VibeProxy-arm64.dmg: OK
  1. Double-click the DMG and drag VibeProxy.app into /Applications
  2. Eject the DMG and launch VibeProxy from Applications

DMGs are notarized and stapled since v1.8.312 (2026-09-26), so Gatekeeper should stay quiet.

Install with ZIP

  1. Download VibeProxy-arm64.zip (or VibeProxy-x86_64.zip)
  2. Verify the checksum:
bash
shasum -a 256 -c VibeProxy-arm64.zip.sha256
# expect: VibeProxy-arm64.zip: OK
  1. Unzip and move the app:
bash
unzip VibeProxy-arm64.zip
mv VibeProxy.app /Applications/
  1. Launch VibeProxy from /Applications
Download VibeProxy

Verify: VibeProxy.app sits in /Applications and opens without a Gatekeeper dialog. If broken: a “damaged” dialog can still appear on rare occasions. Clear quarantine and retry:

bash
xattr -cr /Applications/VibeProxy.app

First launch

  1. Open VibeProxy from /Applications
  2. A menu bar icon appears (green dot = server running)
  3. Click the icon and select Open Settings
  4. The proxy starts automatically on port 8317

Verify: menu bar status is green, and a quick probe answers on the proxy port:

bash
curl -s -o /dev/null -w "%{http_code}\n" http://localhost:8317
# any HTTP status means the listener is up (404/401 are fine for a bare probe)

If broken: no menu bar icon means an older build with a TCC/entitlements bug (fixed in v1.8.313). Update the app from the releases page.

Connecting your AI subscriptions

With VibeProxy running, authenticate each provider you want to use. The flow is the same everywhere: click Connect in Settings, finish OAuth in the browser, and VibeProxy detects completion automatically.

Connecting Claude Code

  1. Click the VibeProxy menu bar icon and select Open Settings
  2. Click Connect next to Claude Code
  3. Log in when the browser opens
  4. VibeProxy detects completion and stores the tokens

If broken: token expired or provider shows disconnected, use Disconnect then Connect again (see Troubleshooting).

Connecting OpenAI Codex

  1. Click Connect next to Codex in Settings
  2. Complete the browser authentication with your ChatGPT account
  3. Wait for VibeProxy to confirm the connection

Connecting Antigravity (Gemini 3 & Claude-via-Antigravity)

Antigravity vs Gemini CLI

  • Antigravity gives you Gemini 3.x (gemini-3.1-pro-preview, gemini-3-pro-preview) and Claude-via-Antigravity models with a generous shared quota
  • Gemini CLI gives you the 2.x family (gemini-2.5-pro, gemini-2.5-flash)
  • Connect both if you want the full range
  1. Click Connect next to Antigravity
  2. Sign in with your Google account and grant model access
  3. Restart VibeProxy to activate the Gemini 3.x model list

Connecting Gemini CLI

  1. Click Connect next to Gemini
  2. Sign in with your Google account
  3. Select a Google Cloud project (or accept the default)
  4. VibeProxy saves the credentials in ~/.cli-proxy-api/

Connecting GitHub Copilot

  1. Click Connect next to GitHub Copilot in Settings
  2. Enter the device code at GitHub (the code is copied with its hyphen, so a paste just works)
  3. Authorize VibeProxy against your GitHub account

You need an active Copilot subscription. The payoff is worth it: you get Claude, GPT, and Gemini models through GitHub’s API on a separate quota. If you are choosing or rethinking that plan, read our GitHub Copilot Pro guide first.

Copilot premium-request multipliers

Copilot bills premium models at different rates. As of late 2026: claude-opus-4.6 counts as 3x, claude-sonnet-4.6 as 1x, claude-haiku-4.5 as 0.33x. Opus through Copilot is the fastest way to burn a plan. If your usage just went metered, GitHub Copilot usage-based pricing changes and cheaper alternatives walks through the numbers and the exits.

Connecting Kimi, Qwen, and Z.AI GLM

  • Kimi and Qwen: same Connect-and-OAuth flow. Kimi’s model IDs are not enumerated in the official docs; use the provider list inside the app.
  • Z.AI GLM is the odd one out. It takes an API key instead of OAuth: grab one from your Z.AI account (API key management page), then Add Account in VibeProxy Settings and paste it. The key is stored in ~/.cli-proxy-api/ like the OAuth tokens.

Verify (all providers): every provider you intend to use shows Connected in Settings. Then run a test prompt from any configured tool and confirm you get a completion back. An error like auth_unavailable: no auth available means that model’s provider is not authenticated yet.

Setting up Factory AI Droids

Factory Droid CLI is an AI coding agent that takes custom model definitions. Point those definitions at VibeProxy and your subscriptions run the whole session.

Install Factory CLI

bash
curl -fsSL https://app.factory.ai/cli | sh

Configure custom models

Create or edit ~/.factory/config.json. The pattern: Anthropic-shaped Claude models use base_url: http://localhost:8317 with provider: "anthropic"; everything else goes through http://localhost:8317/v1 with provider: "openai". The api_key is a placeholder; VibeProxy ignores it.

json
{
  "custom_models": [
    {
      "model_display_name": "CC: Opus 4.7 (High)",
      "model": "claude-opus-4-7-thinking-32000",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "CC: Sonnet 4.6",
      "model": "claude-sonnet-4-6",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "GPT-5.3 Codex",
      "model": "gpt-5.3-codex",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "Gemini 3.1 Pro",
      "model": "gemini-3.1-pro-preview",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    }
  ]
}

Use Factory with VibeProxy

  1. Launch Factory CLI:
bash
droid
  1. Select a model with /model and pick one of your custom entries
  2. Code. All requests route through VibeProxy automatically.

Verify: /model lists your custom models and a test prompt returns a completion. If broken: 404s from Factory usually mean VibeProxy is not running or the base_url is wrong. Check the menu bar status and the base_url values above.

Full Factory configuration with all models

A fuller set covering thinking budgets, Copilot entries, and GLM:

json
{
  "custom_models": [
    {
      "model_display_name": "CC: Opus 4.7 (High)",
      "model": "claude-opus-4-7-thinking-32000",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "CC: Opus 4.7 (Medium)",
      "model": "claude-opus-4-7-thinking-10000",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "CC: Opus 4.7 (Low)",
      "model": "claude-opus-4-7-thinking-4000",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "CC: Opus 4.7",
      "model": "claude-opus-4-7",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "CC: Sonnet 4.6 (High)",
      "model": "claude-sonnet-4-6-thinking-32000",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "CC: Sonnet 4.6",
      "model": "claude-sonnet-4-6",
      "base_url": "http://localhost:8317",
      "api_key": "dummy-not-used",
      "provider": "anthropic"
    },
    {
      "model_display_name": "GPT-5.5 (High)",
      "model": "gpt-5.5(high)",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GPT-5.5",
      "model": "gpt-5.5",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GPT-5.5 (Low)",
      "model": "gpt-5.5(low)",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GPT-5.3 Codex (High)",
      "model": "gpt-5.3-codex(high)",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GPT-5.3 Codex",
      "model": "gpt-5.3-codex",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "Gemini 3.1 Pro",
      "model": "gemini-3.1-pro-preview",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "Gemini 3 Pro",
      "model": "gemini-3-pro-preview",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GH: Opus 4.6 (via Copilot)",
      "model": "claude-opus-4.6",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GH: Sonnet 4.6 (via Copilot)",
      "model": "claude-sonnet-4.6",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GLM-5",
      "model": "glm-5",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "GLM-4.7",
      "model": "glm-4.7",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    },
    {
      "model_display_name": "Qwen3 Coder Plus",
      "model": "qwen3-coder-plus",
      "base_url": "http://localhost:8317/v1",
      "api_key": "dummy-not-used",
      "provider": "openai"
    }
  ]
}

Setting up Amp Code

Amp Code by Sourcegraph works with VibeProxy, but the rules changed. The old secrets.json hack is gone, and there is no fallback to Amp credits for models you have not authenticated. That matches why Amp Code’s free tier was paused: when the provider toggles are all off, Amp falls back to its own rate-limited free tier. For anything else, you authenticate the provider or you get nothing.

No fallback anymore

If Amp requests a model whose provider is not connected, you get auth_unavailable: no auth available. Amp does not silently spend its own credits. Authenticate the provider (or pick a model from a provider you did authenticate).

Point Amp at VibeProxy

bash
mkdir -p ~/.config/amp
echo '{"amp.url": "http://localhost:8317"}' > ~/.config/amp/settings.json

Authenticate the providers you want

Use the bundled backend binary inside the app bundle (the cli-proxy-api-plus filename is a legacy name):

bash
VP=/Applications/VibeProxy.app/Contents/Resources
$VP/cli-proxy-api-plus -config $VP/config.yaml -claude-login   # Claude Code
$VP/cli-proxy-api-plus -config $VP/config.yaml -codex-login    # ChatGPT / Codex
$VP/cli-proxy-api-plus -config $VP/config.yaml -login          # Gemini

You can also run amp login now: it redirects directly to ampcode.com (the old “403 Invalid state parameter” failure was fixed in v1.8.51). Restart VibeProxy after the logins.

How Amp traffic splits: management calls (/auth/cli-login, /api/*) forward to ampcode.com; model calls (/provider/*) go to the CLIProxyAPI backend on 8318. Nothing to configure for that split.

Use Amp with VibeProxy

bash
# Interactive mode
amp

# Direct prompt
amp "Say hello"

Verify: amp "Say hello" returns a greeting. If broken: auth_unavailable means the model’s provider has no login. Re-run the matching login command above and restart VibeProxy.

Use VibeProxy with any AI coding tool (no API keys needed)

Any tool that accepts a custom OpenAI-compatible (or Anthropic-compatible) endpoint can use VibeProxy. The generic recipe:

Setting Value
Base URL http://localhost:8317/v1 (or http://localhost:8317 for Anthropic-shaped APIs)
API Key dummy or any placeholder string
Model a current model ID from the reference below

If you enabled the LAN toggle (see Security), substitute the machine’s LAN IP for 127.0.0.1/localhost in these URLs. Editors like Windsurf that accept custom OpenAI-compatible endpoints follow the same table.

Verify: one test prompt per configured tool. Any completion proves the path works.

Zed editor

Zed is a fast code editor with AI features. In its AI settings (open Settings with Cmd+,), set the API endpoint to http://localhost:8317/v1 and use dummy as the API key. If the settings UI has moved fields around since Oct 2026, look for the custom provider / custom endpoint section.

OpenCode

For OpenCode and similar CLI tools, the environment-variable pattern usually works:

bash
export OPENAI_API_BASE="http://localhost:8317/v1"
export OPENAI_API_KEY="dummy"

OpenCode’s config surface keeps moving; if that fails, check our OpenCode setup guide for the current config file shape. If you would rather buy cheap model access than proxy subscriptions, OpenCode has an OpenCode Go plan with a pool of low-cost models.

VS Code extensions (Roo Code, Continue, etc.)

Many VS Code extensions support custom endpoints:

json
{
  "provider": "OpenAI Compatible",
  "base_url": "http://localhost:8317/v1",
  "api_key": "dummy-not-used",
  "model": "claude-sonnet-4-6"
}

Any other tool

Same three settings from the table above. The Anthropic-native path (Claude-shaped SDKs) uses http://localhost:8317 as the base with no /v1; the OpenAI-shaped path uses /v1. If you are sandboxing third-party CLIs, run AI CLI tools in Docker or Podman safely and point them at the host’s loopback address.

Use VibeProxy with Codex CLI itself

You can even point Codex CLI at VibeProxy instead of at ChatGPT directly. Edit ~/.codex/config.toml:

Option 1: Simple base_url override (recommended)

toml
base_url = "http://127.0.0.1:8317/v1"

This is the option to use. It keeps Codex’s native ChatGPT surfaces (Quick Chat, “More details”, Computer history) working.

Option 2: Explicit provider block

toml
model_provider = "cliproxyapi"

[model_providers.cliproxyapi]
name = "cliproxyapi"
base_url = "http://127.0.0.1:8317/v1"
wire_api = "responses"

Upstream CLIProxyAPI docs recommend this shape, but there is a catch: see the warning below.

Quick Chat breaks with the explicit provider block (issue #544)

With the explicit model_provider block, Codex’s built-in ChatGPT surfaces (Quick Chat, “More details”, Computer history) consistently fail. Tracked as issue #544 in the VibeProxy repo. The simple base_url override keeps them working, which is why I recommend it for most setups.

Verify: run codex and confirm it picks up the proxy’s model list. For the related Codex desktop app, see how to use the Codex app with any model.

Multi-account & provider priority

Two features in the same Settings panel do the heavy lifting for day-to-day use.

Multi-account (v1.7.0+)

Add several accounts per provider. VibeProxy distributes requests across them round-robin and fails over automatically when one account hits a rate limit. Account lists are collapsible; accounts can be removed individually.

[COST] This is the practical rate-limit mitigation for Max/Pro users. Two Claude accounts means one can cool off in the 5-hour window while the other serves. It does not double your per-window quota; it spreads and backstops it.

Provider priority toggles (v1.8.51+)

Enable or disable providers in Settings. Disabled providers have their models excluded from the served list (oauth-excluded-models under the hood), and the change hot-reloads without restarting the app.

When to use the toggles:

  • Pin a single provider when you want predictable routing
  • Skip a rate-limited provider without logging it out
  • Resolve model-name collisions between providers (the dot-vs-dash problem in the models reference)

Verify: disable a provider in Settings and confirm its models disappear from your connected tool’s model list without a restart.

Does adding a second account double my quota?

No. Providers meter per account (and often per organization). A second account gives you a second independent window and failover capacity, not a merged super-quota. If both accounts are on the same team plan, they may share team-level limits.

VibeProxy supported models reference (2026)

Model IDs below are current as of 2026-10-21 (v1.8.341, from the bundled FACTORY_SETUP docs). VibeProxy ships near-daily and model IDs follow the providers, so if a listed ID returns not-found, check Settings, run copilot --help for Copilot names, or read the bundled setup docs.

Dot vs Dash: Two Different Backends

This is the most common footgun. Copilot models use dot notation (claude-opus-4.6, claude-sonnet-4.6) and route through provider: "openai" with the /v1 base URL. Claude Code subscription models use dash notation (claude-opus-4-7, claude-sonnet-4-6) and route through provider: "anthropic" with the bare base URL. Mixing the two fails with model-not-found errors.

Claude models (via Claude Code subscription)

Model ID Notes
claude-opus-4-7 Strongest current Opus
claude-sonnet-4-6 Default coding model
claude-opus-4-7-thinking-32000 Opus with extended thinking (32K budget)
claude-opus-4-7-thinking-10000 Opus, medium thinking
claude-opus-4-7-thinking-4000 Opus, low thinking
claude-sonnet-4-6-thinking-32000 Sonnet with extended thinking

Date suffixes are gone. claude-opus-4-5-20251101 and friends no longer appear; use the clean IDs above.

OpenAI GPT models (via ChatGPT subscription)

All through base_url: http://localhost:8317/v1:

Model ID Notes
gpt-5.5 Current flagship
gpt-5.5(high) Higher effort variant
gpt-5.5(low) Cheaper/faster variant
gpt-5.4 Previous flagship
gpt-5.4(high) Higher effort
gpt-5.4-mini Small model
gpt-5.3-codex Coding-tuned
gpt-5.3-codex(high) Coding, high effort
gpt-5.3-codex-spark Lightweight coding

Gemini models (via Antigravity & Gemini CLI)

Model ID Notes Auth
gemini-3.1-pro-preview Current Gemini 3.1 Antigravity
gemini-3.1-flash-image-preview Image generation Antigravity
gemini-3-pro-preview Gemini 3 Pro Antigravity
gemini-3-pro-image-preview Gemini 3 Pro with vision Antigravity
gemini-claude-opus-4-6-thinking Claude Opus via Antigravity (generous quota) Antigravity
gemini-claude-sonnet-4-6-thinking Claude Sonnet thinking via Antigravity Antigravity
gemini-claude-sonnet-4-6 Claude Sonnet via Antigravity Antigravity
gemini-2.5-pro Gemini 2.5 Pro Gemini CLI
gemini-2.5-flash Gemini 2.5 Flash Gemini CLI
gemini-2.5-flash-lite Gemini 2.5 Flash Lite Gemini CLI

The Claude-via-Antigravity IDs are worth knowing: they run Claude models on a quota shared with Sonnet and GPT-OSS rather than your Claude subscription.

GitHub Copilot models (dot notation: claude-opus-4.6)

Model ID Premium rate
claude-opus-4.6 3x
claude-sonnet-4.6 1x (Copilot default)
claude-haiku-4.5 0.33x
gpt-5.3-codex 1x
gemini-3.1-pro-preview 1x

All Copilot entries go through provider: "openai" + /v1. Official tip: run copilot --help to discover the current Copilot model names when something 404s.

Qwen, Kimi & Z.AI GLM models

Provider Model IDs Notes
Qwen qwen3-coder-plus, qwen3-coder-flash OAuth; see also Qwen 3.6 models for AI coding agents
Z.AI GLM glm-5, glm-4.7, glm-4-plus, glm-4-air, glm-4-flash API key; cheap coding models, featured in the cheapest AI models for coding agents roundup
Kimi see in-app provider list OAuth; official docs do not enumerate model IDs yet

[COST] GLM is the budget lane here: an API key from Z.AI GLM plus glm-4.7 or glm-5 handles grunt work without touching your Claude quota. Keep your Opus/Claude budget for the hard problems.

Extended thinking mode

VibeProxy exposes Claude’s Extended Thinking by encoding the budget in the model name. Append a suffix to any Claude model ID:

text
{model-name}-thinking-{NUMBER}

Presets (same labels the Claude ecosystem uses):

  • -thinking-4000: “think” (~4K tokens)
  • -thinking-10000: “think harder” (~10K tokens)
  • -thinking-32000: “ultra think” (~32K tokens)

Works on Claude 3.7+, Opus 4/4.7, and Sonnet 4/4.6. Example:

json
{
  "model": "claude-opus-4-7-thinking-10000"
}

What happens: VibeProxy strips the -thinking-10000 suffix, sets the thinking budget parameter to 10,000 tokens, and forwards the request. The response includes the reasoning steps.

Advanced config.yaml overrides

Power users get a persistent override file at ~/.cli-proxy-api/config.yaml. VibeProxy merges it with the bundled defaults into ~/.cli-proxy-api/merged-config.yaml and hot-reloads without an app restart.

Never edit merged-config.yaml

merged-config.yaml is regenerated on every merge. It is a read-only diagnostic: useful for confirming what the effective config is. Put your edits in ~/.cli-proxy-api/config.yaml.

The example from the README worth knowing: force low verbosity on all GPT requests.

yaml
payload:
  default:
    - models:
        - name: "gpt-*"
      protocol: "codex"
      params:
        "text.verbosity": "low"

Two semantics to choose from:

Key Behavior
default Fills the parameter only when the client sends nothing (won’t override a tool that sets its own value)
override Always enforces the value, regardless of what the client sent

Verbosity values: low, medium, high.

Verify: change the file and send the next request. The change should apply without restarting VibeProxy. If broken: a YAML typo means your rule silently does not apply; check the app/proxy log, then diff against merged-config.yaml to see whether your rule made it into the effective config.

Troubleshooting

VibeProxy menu bar status
  • A green dot means the server is running, a red dot means it stopped
  • Click the status to toggle the server on and off

No icon at all? That was a TCC/entitlements bug in older builds, fixed in v1.8.313 (2026-09-27). Update the app.

Connection issues
Problem Solution
Can’t connect to Claude/Codex Re-click Connect in VibeProxy settings
Factory shows 404 errors VibeProxy is not running or base_url is wrong (check menu bar)
Authentication expired Disconnect and reconnect the provider
Port 8317 or 8318 already in use Quit other VibeProxy or CLIProxyAPI instances. Orphaned processes sometimes survive app quits: lsof -i :8317 -i :8318, then kill the PID
auth_unavailable: no auth available That model’s provider has no login. Authenticate it (see the Amp login commands above) and restart VibeProxy
Login buttons fail with “flag provided but not defined” App/backend version drift, fixed in v1.8.313. Update VibeProxy
Gemini returns 401 Verify the Google Cloud project has the Gemini API enabled
Model not found Check dot-vs-dash notation (Copilot vs Claude Code). For Copilot names, run copilot --help
Usage drains faster than expected

Community-reported (issues #313 and #292, both closed): the same prompt can burn subscription usage 2-3x faster through third-party CLIs like Droids than through native Claude Code or Codex. That figure comes from user reports, not a controlled benchmark.

What I would do about it:

  • Watch the usage dashboards on the provider side, not just the tool’s token counter
  • Expect faster drain of the 5-hour window during heavy agentic sessions
  • Use smaller or thinking-off models for grunt work; save Opus/thinking for the hard tasks
  • Spread load across providers and accounts (see Multi-Account & Provider Priority)
Enterprise / Team Claude accounts

If your dashboard shows plan: free or 5-hour-limit 0% on a paid enterprise account: the claude.ai account endpoint returns a Cloudflare challenge instead of JSON for organization accounts, so the proxy cannot read your plan or usage and fails to track or fail over. Tracked as issue #382, open since 2026-05-27. If you rely on usage tracking or auto-failover, test this on your org account before depending on it.

'App is damaged' error

Remove quarantine attributes and reopen:

bash
xattr -cr /Applications/VibeProxy.app

Rare since notarized and stapled DMGs shipped in v1.8.312, but the fix is still the right one.

Verification checklist
  • VibeProxy is running (menu bar icon shows green)
  • Every provider you need shows Connected in Settings
  • Your tool’s base URL points at localhost:8317 (with /v1 for OpenAI-shaped APIs)
  • API key in the tool is set (any placeholder works)
  • A test prompt (“what day is it?”) returns a completion
Uninstall / rollback
  1. Quit VibeProxy (menu bar icon, or pkill -f VibeProxy)
  2. Delete /Applications/VibeProxy.app
  3. Optionally remove ~/.cli-proxy-api/. This deletes the OAuth tokens and is the real “disconnect all providers” step
  4. Revert each tool’s config: ~/.factory/config.json, ~/.config/amp/settings.json, ~/.codex/config.toml, editor endpoint settings
  5. Check for orphaned listeners: lsof -i :8317 -i :8318 and kill any leftovers

Warning: removing ~/.cli-proxy-api/ logs you out of every proxied account. You will need to re-run the OAuth flows after a reinstall.

Security considerations

Important security information

  • OAuth tokens are stored locally in ~/.cli-proxy-api/ with 0600 permissions
  • Since v1.8.313 (2026-09-27) port 8317 binds to 127.0.0.1 only. Earlier builds bound all interfaces; the current behavior is enforced
  • The “Allow LAN connections (port 8317)” toggle in Settings exposes the proxy to your network, with a warning in the UI and bind-mode diagnostics in the proxy log. If you enable it, any machine on the LAN can spend your subscription quota. Use the machine’s LAN IP in tool configs and treat the toggle as temporary
  • Traffic to AI providers is TLS end to end
  • Tokens refresh before expiry, so you are not storing raw passwords

Terms of service risk

Using subscription OAuth tokens through a third-party proxy may violate provider terms of service. The project’s own docs warn about account suspension and permanent bans. VibeProxy adds an official escape hatch: you can route Claude requests through Vercel’s AI Gateway for sanctioned access to a Claude Max subscription instead of calling Anthropic with OAuth tokens directly. Vercel’s AI Gateway is a paid, credit-based service; see Vercel’s AI Gateway docs for current pricing and setup. Use at your own risk.

Use cases and benefits

For professional developers

  • Your existing flat-rate subscription covers tools that would otherwise bill per token
  • Switch between Claude, GPT, Gemini, Copilot-hosted models, and GLM without juggling separate API keys
  • Run Factory, Amp, and Zed against the same proxy, and add a second account for failover
  • Extended thinking gives you reasoning budgets for architecture decisions and hard debugging
  • When one terminal is not enough, run AI coding agents on macOS with cmux

For hobbyists and learners

  • A single subscription goes further when every tool can spend its quota
  • Try different coding tools without opening a new API bill
  • Watching traffic hit a local proxy teaches you how AI APIs work
  • Side projects can run on premium models within your plan’s quota

[COST] The win to remember: avoid per-token API bills for work your flat-rate subscription already covers. When a project genuinely needs metered scale, switch to API keys deliberately instead of leaking usage through a proxy.

For more AI programming guidance:

Frequently asked questions

Is VibeProxy free?

Yes. MIT-licensed, $0, download it from GitHub and run it as long as you like. You still need at least one paid AI subscription (or a GLM API key) for the proxy to have anything to route.

Does it work on Intel Macs?

Yes, with caveats. x86_64 builds ship since v1.8.58 (2026-01-15), but the project marks them untested. If you run one, report results upstream so that changes.

Will this get my account banned?

There is real risk. Using subscription OAuth tokens through a proxy can violate provider TOS, and the project’s own docs warn about suspension and permanent bans. If you want lower-risk Claude Max access, use the Vercel AI Gateway route described in Security Considerations.

Can I use it with Linux or Windows?

The VibeProxy app is macOS-only. The underlying CLIProxyAPI runs on other platforms if you start it manually. See the CLIProxyAPI repository.

How do I update VibeProxy?

Since v1.6, VibeProxy checks for updates daily and installs them via Sparkle. Releases are near-daily and often track upstream CLIProxyAPI bumps, so let auto-update do its job. Manual downloads from the releases page work too.

What if a model isn't working?
  1. Confirm the model’s provider is authenticated in Settings
  2. Check the model ID against the reference tables above (watch dot-vs-dash notation)
  3. For Copilot model names, run copilot --help
  4. Restart VibeProxy after config changes (except Priority toggles, which hot-reload)
  5. Check Console.app logs, search for “VibeProxy”
Why does my subscription usage drain faster through VibeProxy?

Community-reported (issues #313/#292): third-party agentic CLIs can burn the 5-hour window 2-3x faster than native Claude Code or Codex for the same nominal work. That number is from user reports, not a measurement. Watch the provider’s own usage dashboard and use cheaper models for routine edits. Details in Troubleshooting.

Do I need API keys?

No. That is the point. Providers are authenticated with OAuth in the browser. The only real API key in the flow is Z.AI GLM’s. Tool-side “API key” fields take a dummy placeholder like dummy-not-used.

Conclusion

VibeProxy lets you use the AI subscriptions you already pay for across every coding tool you own, instead of buying API access a second time. Claude Code, ChatGPT Plus, Gemini, GitHub Copilot, Kimi, Qwen, and Z.AI GLM all route through one local proxy. The tool side is covered too: Factory Droids, Amp Code, Zed, Codex CLI, VS Code extensions, anything OpenAI-compatible. Model IDs stay current with the bundled docs (Opus 4.7, Sonnet 4.6, GPT-5.5, Gemini 3.1, Copilot-hosted models, GLM-5), and the day-to-day controls are there when you need them: multi-account round-robin, Priority toggles, config.yaml overrides, extended thinking. It is a native macOS app, loopback-only by default, with auto-updates and SHA-256 verified releases.

Whether you use GitHub Copilot, Amp Code, Factory Droids, or something else, VibeProxy is the boring local proxy that makes one subscription work everywhere.

Download VibeProxy

Start with the DMG install and the SHA-256 verify step, connect one provider, and run a test prompt before wiring up the rest of your tools.

Did this guide work for you?

Need help with Astro, VPS, or a self-hosted app? View services →