Post

Gemini Claude Code

Posted
By Alex Mathai
6 min read
Gemini Claude Code

Using Gemini Models Inside Claude Code via LiteLLM

How to route Claude Code’s CLI through a LiteLLM proxy so it talks to Vertex AI (Gemini) on the backend — with both Google AI Studio and Vertex AI options.

1. Why Would You Do This?

Claude Code is a solid agentic coding CLI. But sometimes you want to run a different model behind it:

  • Stay within your GCP billing instead of paying Anthropic separately
  • Compare model outputs on the same coding tasks
  • Use a different model’s context window for specific workloads

The trick is that Claude Code speaks the Anthropic Messages API. Gemini speaks Google’s API. You need a translator in the middle. That’s where LiteLLM comes in.

2. The Architecture

1
2
3
4
5
6
7
8
9

You (terminal)
    ↓
Claude Code CLI
    ↓  (Anthropic Messages API format)
LiteLLM Proxy (localhost:4000)
    ↓  (translates to Google's API format)
Vertex AI / Google AI Studio
    ↓
Gemini model (3.1 Pro, 2.5 Flash, etc.)

LiteLLM is an open-source proxy that accepts requests in one API format and forwards them to 100+ LLM providers. Here, we configure it to accept Anthropic-format requests from Claude Code and translate them into Vertex AI calls to Gemini.

3. Prerequisites

  • Claude Code installed (docs)
  • Python 3.10+ with pip
  • One of:
    • Google AI Studio API key (Option A — simpler, free tier available)
    • GCP project with Vertex AI enabled (Option B — uses your existing billing)

4. Install LiteLLM

1

pip install 'litellm[proxy]'

This installs both the LiteLLM library and its proxy server.

5. Choose Your Provider

You have two options for how Gemini receives requests. Pick one.

Option A: Google AI Studio (Simpler)

Best for: quick setup, personal use, free tier.

  1. Get an API key at aistudio.google.com/apikey
  2. Export it:
1

export GEMINI_API_KEY="your-key-here"

Option B: Vertex AI (GCP)

Best for: existing GCP projects, enterprise billing, higher rate limits.

  1. Authenticate with GCP:
1

gcloud auth application-default login
  1. Set your project:
1

export GOOGLE_CLOUD_PROJECT="your-gcp-project-id"

No API key needed — Vertex AI uses Application Default Credentials (ADC).

6. Create the LiteLLM Config

Create a file at ~/.litellm/config.yml:

Option A config (Google AI Studio):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

litellm_settings:
  drop_params: true
  set_verbose: true

model_list:
  # Primary model
  - model_name: gemini-3.1-pro-preview
    litellm_params:
      model: gemini/gemini-3.1-pro-preview
      api_key: os.environ/GEMINI_API_KEY

  # Faster/cheaper option
  - model_name: gemini-2.5-flash
    litellm_params:
      model: gemini/gemini-2.5-flash
      api_key: os.environ/GEMINI_API_KEY

  # Budget option
  - model_name: gemini-3.1-flash-lite
    litellm_params:
      model: gemini/gemini-3.1-flash-lite-preview
      api_key: os.environ/GEMINI_API_KEY

Option B config (Vertex AI):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

litellm_settings:
  drop_params: true
  set_verbose: true

model_list:
  - model_name: gemini-3.1-pro-preview
    litellm_params:
      model: vertex_ai/gemini-3.1-pro-preview
      vertex_project: "your-gcp-project-id"
      vertex_location: "global"

  - model_name: gemini-2.5-flash
    litellm_params:
      model: vertex_ai/gemini-2.5-flash
      vertex_project: "your-gcp-project-id"
      vertex_location: "global"

  - model_name: gemini-3-flash-preview
    litellm_params:
      model: vertex_ai/gemini-3-flash-preview
      vertex_project: "your-gcp-project-id"
      vertex_location: "global"

  - model_name: gemini-3.1-flash-lite
    litellm_params:
      model: vertex_ai/gemini-3.1-flash-lite-preview
      vertex_project: "your-gcp-project-id"
      vertex_location: "global"

Key settings explained:

SettingPurpose
drop_params: trueSilently drops parameters that Gemini doesn’t support (e.g., Anthropic-specific fields) instead of erroring
set_verbose: trueEnables detailed logging — useful for debugging, disable once stable
model: vertex_ai/...The vertex_ai/ prefix tells LiteLLM to use the Vertex AI provider
model: gemini/...The gemini/ prefix tells LiteLLM to use the Google AI Studio provider
vertex_location: "global"Routes to Google’s global endpoint; you can also use us-central1, europe-west4, etc.

7. Start the Proxy

Set a master key (this authenticates Claude Code to the proxy):

1
2

export LITELLM_MASTER_KEY="sk-litellm-$(openssl rand -hex 16)"
echo "Your master key: $LITELLM_MASTER_KEY"

Start the proxy:

1

litellm --config ~/.litellm/config.yml

The proxy starts on http://0.0.0.0:4000. You should see it log the available models.

Verify it works (in a new terminal):

1
2
3
4
5
6
7
8

curl -X POST http://0.0.0.0:4000/v1/messages \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.1-pro-preview",
    "max_tokens": 200,
    "messages": [{"role": "user", "content": "Say hello"}]
  }'

If you get a response back, the proxy is working.

8. Connect Claude Code

In a new terminal, point Claude Code at the proxy:

1
2
3

export ANTHROPIC_BASE_URL="http://0.0.0.0:4000"
export ANTHROPIC_AUTH_TOKEN="$LITELLM_MASTER_KEY"
export CLAUDE_CODE_ATTRIBUTION_HEADER=0

Now launch Claude Code with a Gemini model:

1
2
3
4

claude --model gemini-3.1-pro-preview

# Or a lighter model
claude --model gemini-2.5-flash

That’s it. Claude Code’s full toolset — file editing, bash execution, glob/grep, multi-step planning — now runs on Gemini under the hood.

9. Make It Permanent (Optional)

Add aliases to your ~/.zshrc or ~/.bashrc so you can switch between Claude and Gemini effortlessly:

1
2
3
4
5
6
7
8
9
10

# Claude Code with Gemini backend
alias claude-gemini='ANTHROPIC_BASE_URL="http://0.0.0.0:4000" \
  ANTHROPIC_AUTH_TOKEN="your-master-key" \
  CLAUDE_CODE_ATTRIBUTION_HEADER=0 \
  claude --model gemini-3.1-pro-preview'

alias claude-flash='ANTHROPIC_BASE_URL="http://0.0.0.0:4000" \
  ANTHROPIC_AUTH_TOKEN="your-master-key" \
  CLAUDE_CODE_ATTRIBUTION_HEADER=0 \
  claude --model gemini-2.5-flash'

Then you can just type claude-gemini or claude-flash from any terminal.

10. Model Selection Guide

Which Gemini model to use depends on your task:

ModelUse CaseSpeedCost
Gemini 3.1 ProMulti-file refactors, architecture questionsSlowerHigher
Gemini 2.5 FlashGeneral coding, quick editsFastLower
Gemini 3 FlashGeneral purposeFastLower
Gemini 3.1 Flash LiteSimple queries, high-volume tasksFastestLowest

11. Troubleshooting

“Connection refused” errors: Make sure the LiteLLM proxy is running in another terminal. The proxy must stay alive for the duration of your Claude Code session.

“Model not found” errors: Check that the model_name in your config exactly matches what you pass to claude --model. The names are case-sensitive.

Authentication failures (Vertex AI): Run gcloud auth application-default login again. ADC tokens expire — if you haven’t authenticated recently, refresh them.

Unsupported parameter warnings: This is normal. Claude Code sends some Anthropic-specific parameters that Gemini doesn’t understand. The drop_params: true setting handles this gracefully, but you may still see warnings in verbose mode.

Slow first request: The first request through the proxy takes longer because LiteLLM needs to establish the connection. Subsequent requests are faster.

12. What Works and What Doesn’t

Claude Code’s tooling layer is model-agnostic — it handles tool calls, file operations, and shell execution regardless of which model generates the instructions. In practice:

Works well:

  • File reading, editing, and creation
  • Bash command execution
  • Glob and grep searches
  • Multi-step task planning
  • Code generation and refactoring

May degrade:

  • Complex tool-use chains (Gemini occasionally formats tool calls differently than Claude expects)
  • Very long conversations (token counting differences between providers)
  • Some Claude-specific features like extended thinking

The experience works for most coding tasks. The main trade-off is that you lose native Anthropic API compatibility in exchange for provider flexibility.

The LiteLLM proxy approach works with any provider LiteLLM supports — not just Gemini. The same pattern applies to OpenAI, Mistral, Deepseek, or any other model you want to run through Claude Code’s interface.

Technical
This post is licensed under CC BY 4.0 by the author.