Sharp AI Hub
Subscribe

How to Run Claude Code Without Paying for Anthropic's API

AI

Claude Code is one of the best terminal agents out there. Using it isn't the problem — the bill is, if you don't have a Claude subscription and don't want to put a card on Anthropic's API.

There's a clean way out: Claude Code speaks one protocol (the Anthropic Messages API), and anything that speaks it can sit underneath as the engine. Swap Anthropic for a free model, a cheaper provider, or a local model on your laptop — without touching Claude Code itself.

The short path: 4 environment variables

Claude Code reads the endpoint and key from environment variables. Point them somewhere else:

``` export ANTHROPIC_BASE_URL="https://your-gateway" export ANTHROPIC_AUTH_TOKEN="your-gateway-key" export ANTHROPIC_MODEL="full-model-id" export ANTHROPIC_DEFAULT_HAIKU_MODEL="full-fast-model-id" ```

With that, every Claude Code call goes wherever you send it — OpenRouter, DeepSeek, Gemini, Kimi K3, whatever you want. The `sonnet` alias only works behind gateways that proxy Anthropic; on other models, use the full ID.

The router path: `claude-code-router`

If you want to switch models per task, `claude-code-router` (by `musistudio`, 36k GitHub stars, MIT, now on v3) is a local gateway that intercepts Claude Code's requests and re-routes them to OpenAI, Gemini, DeepSeek, OpenRouter, Mistral, Moonshot and more. It runs on your machine, it's free, and the rules are yours: a strong model to write code, a cheap one for the rest.

Local model, zero accounts

Since January 2026 Ollama serves the Anthropic Messages API directly. Install Ollama, pull a model, point `ANTHROPIC_BASE_URL` at `localhost` — and Claude Code runs offline, with no key at all. LM Studio does the same.

The catch — what nobody puts in the caption

Three things, and they're real:

  1. Weak model, weak agent. Claude Code leans hard on tool calling — editing files, running commands. The model's benchmark score matters less than its tool-use quality. Frontier models hold up; weak ones fumble edits and shell commands, and you spend more time fixing than producing.
  2. It stops being "free" in the subscription sense. The moment you activate a gateway, sessions are billed as API usage by that backend. Your Claude Pro/Max subscription doesn't apply there. "Without paying Anthropic's API" is not the same as "without paying anything" — unless the model is local.
  3. Your code passes through a third party. A cloud gateway sees everything Claude Code reads and writes. Before you route proprietary code through it, verify where the gateway runs and what it retains.

When it's worth it

  • Learning, experimenting, personal projects: router + cheap model, or local Ollama. Worth it now.
  • Work code, on deadlines: a frontier model via API usually costs less in lost time than a weak "free" one.
  • Sensitive code: local model, or nothing.

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →