Setting up Together AI with OpenClaw

I find it frustrating to manage multiple API endpoints just to experiment with different open-source models. Switching between a coding model and a vision model often requires changing authentication logic or base URLs, which interrupts the workflow. Together AI handles this by providing a single OpenAI-compatible API for various models.

What You’ll Need

• A Together AI account
• A TOGETHER_API_KEY
• OpenClaw installed on your system

Quick Start

I recommend storing your API key directly for the Gateway to use. You can get started with this command:

openclaw onboard --auth-choice together-api-key

After that, update your configuration to set a default model. I use this JSON5 structure:

{
  agents: {
    defaults: {
      model: { primary: "together/moonshotai/Kimi-K2.5" },
    },
  },
}

If you prefer to automate the setup without interactive prompts, use the following command:

openclaw onboard --non-interactive \
  --mode local \
  --auth-choice together-api-key \
  --together-api-key "$TOGETHER_API_KEY"

This sets together/moonshotai/Kimi-K2.5 as your default choice.

Troubleshooting

If you run the Gateway as a daemon using launchd or systemd, the process might fail to find your API key. I make sure the TOGETHER_API_KEY is available to the process by adding it to ~/.clawdbot/.env or using env.shellEnv.

What’s Next

You can now use several open-source models through the together provider:

• Llama 3.3 70B Instruct Turbo: Fast instruction following.
• DeepSeek V3.1: Model for coding and reasoning.
• Llama 4 Scout: Vision model for image understanding.
• Kimi K2 Instruct: High-performance model with a 262K context window.

If you need help with other providers, check out the AI Setup Assistant.