Claude Code has rapidly emerged as a powerful, unopinionated command-line tool designed for agentic coding. Developed as an internal research project at Anthropic, it allows developers to integrate Claude directly into their terminal workflows to automate complex tasks like refactoring, testing, and even managing git operations. While natively built for Anthropic’s Claude 3.5 and 4.0 series models, many power users are discovering that the tool's flexibility allows for the integration of alternative LLM models through an LLM gateway or API proxy.
This guide explores how to use a different llm with claude code to break free from the default configuration and leverage a variety of models ranging from OpenAI's GPT series to local models running on your own hardware. By mastering claude code local llm integration, you can create a truly custom development environment.
Claude Code is a REPL (Read-Eval-Print Loop) that acts as an agentic assistant, meaning it doesn't just suggest code; it can execute shell commands, read and write files, and orchestrate subagents to solve complex problems.
By default, Claude Code connects to Anthropic’s official API. It determines which features to enable based on the API format it receives, primarily looking for the Anthropic Messages format (e.g., /v1/messages). It uses these models to reason through a codebase, gather context automatically from files like CLAUDE.md, and use tools via the Model Context Protocol (MCP).
While Claude 3.5 Sonnet is the default for its balance of speed and reasoning, developers often seek alternative LLM models for distinct reasons. Cost Management is a primary driver, as third-party gateways can offer flexible usage-based pricing. Others prioritize Context Management, using different models for background tasks to prevent session clutter. Additionally, Privacy and Local Development needs often lead developers to run a local LLM with Claude Code, ensuring proprietary code never leaves their machine.
For power users, routing Claude Code requests to alternative backends offers distinct functional advantages. One of the most immediate benefits is Speed; smaller models like GPT-4o-mini or Claude 3.5 Haiku can process simple tasks, such as writing git commit messages, much faster than their larger counterparts.
Beyond simple speed, these configurations offer Enhanced Features such as automatic failover and retry logic provided by certain gateways, ensuring coding sessions remain uninterrupted. Experimentation also becomes significantly more accessible; expensive tasks or vibe-coding flow states are more affordable when backed by cheaper or local OSS models like Qwen3 Coder or Mistral-Small. Finally, for organizations, this approach provides Centralized Control, enabling teams to use gateways like LiteLLM for auditing, logging, and budget tracking across multiple engineers.
To use a non-Anthropic model, you must provide Claude Code with an endpoint that mimics the Anthropic API. There are four primary methods to achieve this:
Gateways like OpenRouter, LLMGateway, and ZenMux provide a unified API that translates various model outputs into the Anthropic-compatible format. Using openrouter "anthropic skin" functionality is particularly popular as it allows for a near-seamless drop-in replacement for any claude llm backend with minimal configuration changes.
Tools like LiteLLM or custom Python scripts act as a local proxy between Claude Code and other API providers. This method provides high flexibility and supports in-session model switching across multiple providers, though it requires maintaining a secondary proxy process in the background.
By using Ollama or LM Studio, you can run models directly on your hardware for total privacy and zero per-token costs. Setting up claude code ollama or claude code lm studio is a great way to ensure that your claude llm workflows remain private and secure while avoiding high API costs during intensive coding sessions.
Services like Z.AI or Moonshot AI provide direct, Anthropic-compatible endpoints specifically designed for Claude Code. These are often much cheaper than standard API rates and require no proxy, though users are limited to the specific models hosted by that provider.
The core of any Claude Code LLM configuration involves overriding two environment variables: ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN.
Deploying the "anthropic skin" openrouter feature provides an endpoint that speaks the native protocol directly. This is the fastest way to implement "openrouter" "anthropic skin" routing for your project.
1. Set Environment Variables: Initialize your session by pointing the base URL to OpenRouter and clearing any conflicting keys.
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="your_openrouter_key"
export ANTHROPIC_API_KEY="" # Must be explicitly empty2. Override the Model Tier: Explicitly define which model OpenRouter should route to for the default Sonnet alias.
export ANTHROPIC_DEFAULT_SONNET_MODEL="openai/gpt-5.2-pro"3. Start Claude: Launch the tool with the claude command and verify your connection status using /status.
LiteLLM is ideal for those who want to switch between local and cloud providers dynamically.
1. Installation & Configuration: Install the proxy with pip install 'litellm[proxy]' and create a config.yaml to map your desired models.
model_list:
- model_name: custom-sonnet
litellm_params:
model: openai/gpt-4o
api_key: os.environ/OPENAI_API_KEY2. Launch & Connect: Start the proxy server using litellm --config config.yaml, then point Claude Code to your local endpoint (usually http://0.0.0.0:4000).
For fans of lm studio claude code setups, or those preferring the simplicity of claude code ollama, pointing to a local server is straightforward.
1. Point to Localhost: Direct Claude Code to your running Ollama instance by updating the base URL variable.
export ANTHROPIC_BASE_URL="http://localhost:11434"
export ANTHROPIC_API_KEY="ollama"2. Run with Model Selection: Launch the tool while specifying your local model as the target.
claude --model qwen3-coderTo avoid re-exporting these variables every session, add them to your shell profile (~/.zshrc or ~/.bashrc). For more advanced control, you can use apikeyhelper claude code in your settings.json to handle rotating keys, per-user authentication, or even managing different claude llm profiles across projects.
Using alternative models is a powerful workflow enhancement, but it comes with specific technical constraints. Most critically, Tool-Calling Requirement is non-negotiable; Claude Code relies on agentic behaviors to read files and run terminal commands. If your chosen model lacks native tool-calling support, the session will simply fail.
API Compatibility is another hurdle, as gateways must forward specific headers like anthropic-beta to maintain full functionality. Furthermore, Context Window Limits can be an issue the tool's system prompt alone can exceed 20k tokens, which may overwhelm smaller models. Finally, users should note that MCP Constraints often limit support to HTTP servers in proxy setups, and the general Stability of these unofficial configurations can vary as the Claude Code tool continues to evolve.
Whether this path is right for you depends on your technical needs and budget. Developers on a Budget or those hitting Pro plan limits will find significant value in usage-based gateways. Local-First Advocates with sufficient hardware can unlock unparalleled privacy, while Power Users can leverage multi-provider setups to use specialized models for specific tasks.
However, for Engineers needing maximum reliability, the official Anthropic models remain the most tested and lucky in terms of adhering to complex system prompts. New Users should also likely stick to the defaults, as the learning curve of agentic coding is steep enough without the added complexity of proxy troubleshooting.
Ultimately, the ability to treat Claude Code as an LLM-agnostic digital intern allow you to build more ambitious projects at a fraction of the cost.
Yes. By overriding the ANTHROPIC_BASE_URL, you can route requests to any model provider that supports the Anthropic Messages API format.
Yes, though not natively. You can use tools like Ollama or LM Studio to host a local server and then point Claude Code to that local address using environment variables.
Third-party gateways are not audited by Anthropic. Ensure your gateway provider has a privacy policy that meets your requirements regarding source code logging.
There is no official confirmation. However, the tool is designed to be unopinionated, making it easy for developers to integrate their own backends.
Claude Code’s unopinionated and flexible architecture transforms it from a simple terminal client into a powerful, model-agnostic platform for agentic coding. While natively optimized for Anthropic’s flagship models, its reliance on standard environment variables allows developers to redirect requests to an expansive ecosystem of gateways, proxies, and self-hosted models.
By integrating alternative models, you can achieve a fine-tuned balance between high-reasoning performance and cost-effective execution. Whether leveraging specialized coding models, using OpenRouter for provider diversity, or running local OSS models via Ollama for total privacy, the ability to switch backends ensures your workflow is never gated by a single provider's limits.
Think of Claude Code as a highly capable specialized toolkit. While the manufacturer provides premium power cells, the tool is designed with a universal port. By adapting different batteries from high-capacity cloud models to rechargeable local ones you ensure that your development environment remains powered, efficient, and perfectly tailored to the demands of your project.
Continue exploring the future of GenAI
Master GGUF quantization with Llama.cpp. Expert guide covering Q4/Q5/Q8 formats, I-Quants, Imatrix optimization, Blackwell GPU builds, and speed benchmarks for 2026.
Step-by-step technical guide on integrating GPT-5.2 and the Responses API into modern web frameworks with best practices for security and cost.
Explore the top 30+ uncensored open-source AI models on Hugging Face for 2026. Includes Llama, Mistral, and Qwen variants for local unfiltered inference.
Loading comments...