Claude Code with Other Models: Ollama, OpenRouter and More

Yes, you can use Claude Code with other models. Claude Code sends every request to the address in ANTHROPIC_BASE_URL, so any server that speaks the Anthropic Messages API can answer it, and an LLM gateway can translate for providers that do not. Anthropic documents the variables but does not support non-Claude models, so test your setup on real tasks before you depend on it.

How does Claude Code connect to a different model?

Claude Code reads its endpoint and credential from environment variables. Anthropic's gateway documentation names the ones that matter here.

• ANTHROPIC_BASE_URL is the address of the gateway or model server. Claude Code posts to /v1/messages under it.
• ANTHROPIC_AUTH_TOKEN sends your credential as an Authorization: Bearer header.
• ANTHROPIC_API_KEY sends it as an x-api-key header. Use the one your server expects.

When one of the credential variables is set, Claude Code uses it instead of a saved claude.ai login. Set both the address and the credential. Anthropic's docs warn that with only the address set, a saved claude.ai login stays the active credential. Run /status in a session to confirm: the Anthropic base URL line shows where requests go, and the credential line names the variable in use.

# bash or zsh (values from Ollama's own docs)
export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=""
claude --model qwen3.5

# PowerShell
$env:ANTHROPIC_BASE_URL = "http://localhost:11434"
$env:ANTHROPIC_AUTH_TOKEN = "ollama"
$env:ANTHROPIC_API_KEY = ""
claude --model qwen3.5

Model choice has its own variables. ANTHROPIC_MODEL sets the model for one session. ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL and ANTHROPIC_DEFAULT_HAIKU_MODEL decide what the opus, sonnet and haiku aliases resolve to, and CLAUDE_CODE_SUBAGENT_MODEL sets the default for subagents. You can put any of them in the env block of ~/.claude/settings.json instead of exporting them each time. The official gateway connection guide has the full list.

Do you need a gateway to use Claude Code with a local model?

Often not. Ollama, LM Studio and the llama.cpp server each expose an Anthropic-compatible endpoint, so Claude Code can talk to them directly.

Server	Address to use	What the docs say
Ollama	http://localhost:11434	A subset of the Messages API on /v1/messages. Its compatibility page lists no token counting, no tool_choice and no prompt caching.
LM Studio	http://localhost:1234	Exposes POST /v1/messages. Its Claude Code example sets ANTHROPIC_AUTH_TOKEN=lmstudio.
llama.cpp	The host and port you give llama-server	The server README lists Anthropic Messages API compatible chat completions. OpenAI-style function calling needs the --jinja flag.

OpenRouter also publishes an Anthropic-compatible address, https://openrouter.ai/api. Put your OpenRouter key in ANTHROPIC_AUTH_TOKEN and set ANTHROPIC_API_KEY to an empty string. Its Claude Code guide adds that Claude Code is optimized for Anthropic models and may not work correctly with other providers. For Ollama specifically, Ollama's integration page also documents ollama launch claude, which sets everything up for you.

Start with a direct connection. It is the simplest test of whether a model can drive Claude Code at all. A gateway earns its place when one of these applies:

• The provider only offers OpenAI Chat Completions or Gemini formats. The gateway translates requests, streaming and tool calls.
• You want different models for different jobs, which is tier routing, covered next.
• You want a stand-in when a model errors or hits a rate limit.
• You want provider keys in one place instead of in every tool's config.
• You want a log of every request with its cost.

A gateway is also one more moving part. Anthropic's compatibility guide says Claude Code adds capabilities with each release, and a gateway that does not forward them breaks the matching features, so keep yours updated.

What is tier routing in a Claude Code gateway?

Claude Code does not use one model for everything. The opus alias is for complex reasoning, sonnet for daily coding and haiku for simple tasks. Anthropic's docs also say that pinning the haiku model moves background work such as session titles onto it. A gateway that does tier routing sees which tier Claude Code asked for and sends the request to a model you chose.

In My Claude Code the settings are MODEL (used when nothing more specific matches), MODEL_OPUS, MODEL_SONNET and MODEL_HAIKU, plus MODEL_FABLE and MODEL_MYTHOS for the other two tiers. Each value is a slug in the form <provider-id>/<exact-provider-model-id>. The values below only show the format. Pick real ones from the Models page, which lists what each configured provider publishes.

MODEL=deepseek/deepseek-chat
MODEL_OPUS=anthropic/claude-sonnet-4-6
MODEL_SONNET=mistral/devstral-small-latest
MODEL_HAIKU=ollama/<model-tag>

Tier	Typical job	A sensible choice
Opus	Planning, hard debugging, large refactors	The strongest model you can afford, which can still be a Claude model on an API key
Sonnet	Everyday edits and questions	A mid-priced cloud model
Haiku	Background work and simple lookups	A small, fast model, local if your hardware allows

This is the same idea as the routing in our post on cutting LLM costs by 99%, which sends each request to the cheapest model that passes your checks.

How do fallback chains work?

Each tier gets an ordered list of stand-ins. My Claude Code walks the list when a model errors, rate-limits or refuses. Two related controls keep a bad route from stalling you. A model that answers 429 is benched for a while instead of being retried repeatedly, and a first-token deadline turns a silent model into a failure that moves to the next entry.

A typical chain for the sonnet tier puts a local model first because it costs nothing per token, a cheap cloud model second and a stronger cloud model last. Keep chains short. Claude Code's own docs also say every model has its own prompt cache, so after a failover the stand-in reads the whole conversation without cache hits. Fallbacks protect the session, and the first turn after one can be slower and pricier.

What do you give up with a non-Claude model?

Anthropic's gateway documentation says Anthropic does not support routing Claude Code to non-Claude models through any gateway. It works, and you are on your own when it does not. Expect these differences.

Area	What changes	What to do
Tool calling	Claude Code is built around Anthropic models, and models differ in how reliably they emit tool calls. Ollama lists tool_choice as unsupported.	Choose models that advertise tool use. My Claude Code probes what a model can do instead of trusting the catalogue.
Context length	Claude Code sends a long system prompt and tool definitions every turn. Ollama recommends a context of 64k or higher. For model IDs it does not recognise, Claude Code assumes a 200K window.	Set CLAUDE_CODE_AUTO_COMPACT_WINDOW to your model's real limit. Claude Code clamps it to at least 100,000 tokens, so models with smaller windows are a poor fit.
Prompt caching	Ollama lists prompt caching as unsupported, so a local model reprocesses the whole conversation each turn. Cloud providers vary.	Keep sessions short and run /clear between tasks.
Thinking and effort	Support is partial. Ollama accepts budget_tokens and does not enforce it.	Lower the effort level on small models.
Guarantees	No vendor tests your exact combination.	Run your own checks.

Test before you trust. Pick three tasks from your real work, for example a rename across several files, a failing test to fix and a module to explain. Run each one on every candidate model through the same setup, and keep the model that finishes them with the fewest corrections. Leave a Claude route in the opus tier for the cases that need it. If the reason you are looking at other models is cost, read Claude Code free for the options and their limits.

How do you set it up with My Claude Code?

My Claude Code (MCC) is a local gateway that Danube Labs built and maintains. It translates between four API formats, reaches 57 providers including Ollama, LM Studio and llama.cpp, and runs on your machine. These steps take a few minutes.

1. Install Claude Code. MCC does not install it for you. On macOS, Linux and WSL run curl -fsSL https://claude.ai/install.sh | bash. In Windows PowerShell run irm https://claude.ai/install.ps1 | iex.
2. Install MCC. Run npx @firedmosquito831/my-claude-code to try it, or npm install -g @firedmosquito831/my-claude-code to keep it. uv tool install my-claude-code works if you manage Python 3.14 yourself, and desktop apps exist for Windows, macOS and Linux. Close and reopen your terminal, then check mcc-server --version.
3. Start the server. Run mcc-server and open http://127.0.0.1:8082/admin.
4. Add a provider. On the Providers page click Configure on a provider, paste the key, then Validate and Apply. For Ollama, LM Studio or llama.cpp, enter the base URL (OLLAMA_BASE_URL, LM_STUDIO_BASE_URL or LLAMACPP_BASE_URL) instead of a key.
5. Set the tiers. On Model Config choose MODEL, then MODEL_OPUS, MODEL_SONNET and MODEL_HAIKU, each with a fallback chain.
6. Launch. Run mcc-claude. Use /status to see the base URL, and open the Analytics page to watch each request with its cost.

Your provider keys stay in one local .env file, and MCC has no telemetry. It is dual-licensed, AGPL-3.0-or-later or commercial, and independent of Anthropic. It began as a fork of free-claude-code by Alishahryar1 and became a standalone project on 5 September 2026. Source and docs live on GitHub, and the project page has more detail.

What breaks most often, and how do you fix it?

These rows come from Anthropic's own gateway troubleshooting table.

Symptom	Cause	Fix
401 errors	The credential is in a header the server does not read	Move it between ANTHROPIC_AUTH_TOKEN and ANTHROPIC_API_KEY
400 naming context_management or Extra inputs are not permitted	The upstream rejects fields Claude Code sends to Anthropic-format endpoints	Set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1
400 stating a context or token limit	The server enforces a smaller window than Claude Code assumes	Run /compact and set CLAUDE_CODE_AUTO_COMPACT_WINDOW
Model missing from /model	Gateway discovery keeps only IDs that contain claude or anthropic	Pass the name with --model, or add one entry with ANTHROPIC_CUSTOM_MODEL_OPTION
Login screen even though curl works	No credential variable reached Claude Code before first-run setup	Set ANTHROPIC_AUTH_TOKEN in your shell or in the env block of ~/.claude/settings.json, never in a committed project settings file

Where Danube Labs fits

Danube Labs is a founder-led software, AI and automation studio in Bucharest. We built My Claude Code to manage providers, keys and routing for AI coding agents in one place, and we use the same ideas (routing, evaluation sets, cost tracking) when we build AI features for clients. If you want to choose models for a product with measured quality instead of guesses, see our AI systems service or book a free intro call.