Unlocking Frontier Power: How to Run Amp Using CLIProxyAPI Without Spending Credits

The landscape of AI coding agents is shifting faster than most developers can keep up with. Right now, Amp is sitting at the frontier, offering raw model power that most tools are still trying to figure out. But if you have been using Amp, you know the drill: the “Smart” mode is incredible, but it eats through Amp credits.

If you are already paying for ChatGPT Plus, Claude Pro, or a Google subscription, there is a better way. You can leverage CLIProxyAPI to bridge your existing OAuth subscriptions directly into Amp. This setup effectively allows you to use Amp’s advanced agent features while bypassing the credit consumption by routing requests through your own authenticated quotas.

Here is how to set up this “pro-tier” bridge to get unconstrained token usage with the best models on the market.

Why CLIProxyAPI?

At its core, CLIProxyAPI is a proxy server designed to connect CLI models to an API setup compatible with platforms like OpenAI, Gemini, and Claude. It acts as a universal adapter, allowing you to call these models through standard API requests rather than just terminal commands.

The magic happens with the specialized Amp CLI integration. CLIProxyAPI includes specialized routing that supports Amp’s unique API patterns while maintaining compatibility with standard features. It maps Amp’s specific provider patterns—like /api/provider/{provider}/v1...—to its own internal handlers.

The architecture is simple but powerful:

1. Request Received: Amp sends a request to your local proxy.
2. Local Check: The proxy checks if you have configured an OAuth token for that model locally.
3. Routing: If the model is authenticated (via your subscription), it uses your local OAuth tokens. If not, it falls back to Amp’s backend, which then requires Amp credits.

Step 1: Installing the Infrastructure

Before you can bridge your accounts, you need to get CLIProxyAPI running. You have a few options, but building from source or using the pre-compiled binaries is most common for dev environments.

Building from Source

You will need Go 1.24 or higher.

git clone https://github.com/router-for-me/CLIProxyAPI.git
cd CLIProxyAPI
go build -o cli-proxy-api ./cmd/server

This gives you the cli-proxy-api executable. If you are on macOS, you can also use Homebrew: brew install cliproxyapi.

Step 2: Authenticating Your Subscriptions

This is the most critical step. To use Amp “for free” (using your existing subscriptions), you must authenticate the providers Amp relies on. Amp employs different models for various agent roles:

• Smart Mode: Uses Claude models like Opus and Sonnet 4.5.
• Oracle Subagent: Uses GPT-5 (medium reasoning).
• Librarian/Rush Mode: Uses Claude Sonnet 4.5 or Claude Haiku 4.5.

Run the following commands to log in via OAuth:

• For Google/Gemini: ./cli-proxy-api --login.
• For Antigravity: ./cli-proxy-api -antigravity-login. (Yes, you read it right. We can access Antigravity’s API with the same google account. It’s giving me access to Claude model as well.)
• For ChatGPT Plus/Pro: ./cli-proxy-api --codex-login.
• For Claude Pro/Max: ./cli-proxy-api --claude-login.

If you’re using Github Copilot, you need to use https://github.com/router-for-me/CLIProxyAPIPlus. This one doesn’t have a ready to use build via homebrew so you may need to manually download from release page or just build it yourself.

Once you log in through the browser window that pops up, the system saves your tokens in the ~/.cli-proxy-api directory. This enables the proxy to use your included usage quotas rather than Amp’s backend.

Step 3: Configuring the Proxy for Amp

You need a config.yaml file to tell the proxy how to communicate with Amp’s control plane for management tasks like thread sharing and user authentication.

Create or edit your config.yaml with the following block:

port: 8317
api-keys:
  - "your-secret-key" # You create this key for Amp to talk to the proxy

debug: true
quota-exceeded:
  switch-project: true
  switch-preview-model: true

ampcode:
  upstream-url: "https://ampcode.com"
  restrict-management-to-localhost: true # Extra security for your local machine

Step 4: Pointing Amp to Your Local Proxy

Now that the proxy is ready, you need to tell the Amp CLI to stop talking to the cloud and start talking to localhost. You can do this via environment variables or settings files.

Using Environment Variables

This is the fastest way to test the setup:

export AMP_URL=http://localhost:8317
export AMP_API_KEY=your-secret-key

With these set, the standard amp login command is no longer necessary.

Using Settings Files

For a permanent setup, edit your Amp configuration:

1. URL: In ~/.config/amp/settings.json, set "amp.url": "http://localhost:8317".
2. Key: In ~/.local/share/amp/secrets.json, set "apiKey@http://localhost:8317": "your-secret-key".

Step 5: Validating the Connection

Start your proxy:

./cli-proxy-api --config config.yaml

Now, run a command in Amp:

amp "Analyze the current directory and find all markdown files"

If your proxy logs show requests hitting /api/provider/google/... or /api/provider/openai/... and being handled locally, you are successfully using your own subscription.

The Fallback Safety Net

One of the best features of this integration is the Model Mapping. If Amp requests a model you don’t subscribe to, it will automatically map it to a another model you map to. This ensures Amp never breaks. Here is my mapping.

model-mappings:
  # Model requested by Amp CLI Smart mode
  - from: "claude-opus-4.5"
    to: "gemini-claude-opus-4-5-thinking" # routed to antigravity
  - from: "claude-opus-4-5-20251101"
    to: "gemini-claude-opus-4-5-thinking"

  # Model requested by Amp CLI Rush mode
  - from: "claude-haiku-4-5-20251001"
    to: "gemini-3-flash-preview"

  # model requested by AMP CLI Oracle/Librarian/Review mode
  - from: "gemini-2.5-flash-lite-preview-09-2025"
    to: "gemini-3-flash-preview"
  - from: "gpt-5"
    to: "gemini-3-pro-preview"
  - from: "gpt-5.1"
    to: "gemini-3-pro-preview"

In fact, for the Smart mode, I sometimes run out of Opus access, so I can swap the gemini-claude-opus-4-5-thinking with gemini-claude-sonnet-4-5 or gemini-3-pro-preview instead.

I made a script to automate changing my configuration.

#!/bin/bash

# Script to switch configuration and set model replacement

if [ $# -lt 1 ]; then
    echo "Usage: $0 <model-name> [source-config]"
    echo ""
    echo "Examples:"
    echo "  $0 gemini-claude-opus-4-5-thinking"
    echo "  $0 gemini-claude-sonnet-4-5"
    echo "  $0 gemini-3-pro-preview"
    echo ""
    echo "Optional: specify source config file (defaults to config.yaml.template)"
    exit 1
fi

MODEL="$1"
SOURCE="${2:-config.yaml.template}"
TARGET="config.yaml"

if [ ! -f "$SOURCE" ]; then
    echo "Error: $SOURCE not found"
    exit 1
fi

# Replace all occurrences of "replace_this" with the specified model
sed "s/replace_this/$MODEL/g" "$SOURCE" > "$TARGET"
echo "Switched to model: $MODEL"

Create config.yaml.template file like this

port: 8317
api-keys:
  - "your-secret-key"

debug: true
quota-exceeded:
  switch-project: true
  switch-preview-model: true
ampcode:
  upstream-url: "https://ampcode.com"
  restrict-management-to-localhost: true
  model-mappings:
    # Model requested by Amp CLI Smart mode
    - from: "claude-opus-4.5"
      to: "replace_this" # routed to antigravity
    - from: "claude-opus-4-5-20251101"
      to: "replace_this"

    # Model requested by Amp CLI Rush mode
    - from: "claude-haiku-4-5-20251001"
      to: "gemini-3-flash-preview"

    # model requested by AMP CLI Oracle/Librarian/Review mode
    - from: "gemini-2.5-flash-lite-preview-09-2025"
      to: "gemini-3-flash-preview"
    - from: "gpt-5"
      to: "gemini-3-pro-preview"
    - from: "gpt-5.1"
      to: "gemini-3-pro-preview"

Now i just need to run the following commands:

./switch-config.sh gemini-claude-opus-4-5-thinking
./switch-config.sh gemini-claude-sonnet-4-5
./switch-config.sh gemini-3-pro-preview

It replaces all replace_this placeholders in config.yaml.template with your specified model and writes to config.yaml.

The CLIProxyAPI is already watching the config file and will automatically reload the configuration when it changes.

Advanced Usage: IDE Extensions and Security

The beauty of this setup is that it isn’t limited to the terminal. You can use the same proxy for Amp’s IDE extensions in VS Code, Cursor, or Windsurf. Simply open your IDE settings, set the Amp URL to http://localhost:8317, and provide the local API key you configured.

Security Hardening

Since you are running a proxy that handles sensitive OAuth tokens, security is paramount.

• Localhost Restriction: Enabling restrict-management-to-localhost: true prevents remote access to your management endpoints. This uses the actual TCP connection address to block “drive-by” browser attacks or header spoofing.
• API Key Auth: Ensure api-keys are configured in your config.yaml. As of version 6.6.15, all management routes like /api/auth and /api/threads are protected by this middleware.

Frontier Power, Your Terms

By using CLIProxyAPI as a bridge, you turn Amp into a truly unconstrained tool. You get to keep Amp’s four core principles—unconstrained token usage, always using the best models, raw model power, and a system built to evolve—while leveraging the subscriptions you already pay for.

Whether you are using the Oracle for deep reasoning on a complex bug or spawning subagents to refactor five files at once, this setup ensures that the only thing you are focused on is the code, not the credit balance.