Graphic tutorial: Half an hour to ride your own free big model gateway and let Claude Code/Cursor/Aider all go local/v1图文教程：半小时搭完自己的免费大模型网关，让 Claude Code / Cursor / Aider 全部走本地 /v1

First, make sure that Docker (or Node.js 20 or above, both ways are OK) is installed on the machine. Windows users remember that virtualization and WSL2 are enabled.

Unzip the FreeLLMAPI source code package downloaded from the "AI Tools" section of this site and enter the repository root directory.

Recommended to take the Docker route: Start according to the docker compose configuration in docs/en/install/01-install.md, and the service listens to the local port 3001 by default.

After launching, the browser visits http://localhost: 3001 and can see that the service is normal. The first time you come in is an empty directory because you don't have any keys yet.

Register and generate API Keys from some of the most accessible vendors with free credits: Google (AI Studio), Groq, Cerebras, Mistral, OpenRouter, Cloudflare Workers AI, Cohere, Smart Spectrum Z.ai, NVIDIA NIM. It is recommended to have at least 3 homes, otherwise there is no point in automatically downgrading.

Go back to the Keys page of http://localhost: 3001 and fill in the Keys one by one. After filling in, the router will pull up the list of available models of the vendor, and you will see the number of endpoints rubbing against the increase.

Continue to add more credits if you want: there are more than 30 listed in README, and the official directory page can check the current limit, context window and free token budget of each model.

If you have already run Ollama locally, you can add a custom vendor on the Keys page and point the endpoint to http://localhost: 11434, so that the local model will also enter the routing pool and still have a pocket when the network is completely disconnected.

Open the terminal, execute the configuration command, and point Claude Code to the local service: npx freellmapi setup-claude --url http://localhost: 3001 --api-key The unified key you generated in the panel.

This command will backup your existing configuration file first, and then merge the new base URL with the key, without overwriting the original other settings; you can restore the backup at any time if you are not satisfied.

Other tools have the same approach: Cursor, Cline, Roo Code, Aider, Continue, etc. only need to change the base URL to a local address, replace the Key with a unified key, and the protocol can be kept OpenAI compatible.

If you want to run the Codex CLI, note that it uses the/v1/responses endpoint, and the FreeLLMAPI has been natively implemented; the Gemini CLI can use/v1beta.

First confirm the link with the simplest request: curl http://localhost: 3001/v1/models should return a long list of available models, depending on how many Keys you fill in.

Type the dialog request again, leave the model blank (or fill in auto), let it pick its own; the return value will take the actual manufacturer and which model.

To see the effect of demotion, you can deliberately correct the Key of a vendor, and then make several requests; the log of the panel should automatically cut to the next house after 429 or 5xx, and the request still returns successfully.

Finally, go to the dosage page of the panel to see the statistics: the number of calls and consumption of each Key will be recorded. According to the free quota comparison announced by the manufacturer, you can know how far you are from the explosion quota.

Try Fusion multi-model synthesis: Request the virtual model fusion, and the router will send the same question to multiple free models in parallel, and then the referee model will synthesize it into an answer, which is suitable for judgment questions that need cross-validation.

Solidify the common configuration: Save this set of base URLs as the default configuration in the client, and do not re-match it the next time you open a new project.

Pay attention to two things in long-term maintenance: First, do not use unstable free keys on key production processes; second, pay attention to the terms of each manufacturer, and the free quota is usually prohibited for resale or large-scale commercial use.

If you want the model catalog to always be up-to-date, the paid version of the project ($19/year) can synchronize the real-time catalog on the same day; the free version gets a monthly snapshot, which is generally enough for daily use.