I’ve been running Kimi For Coding as my primary model in OpenClaw (Clawdbot) for a while now. It’s fast, has a massive context window, and supports chain-of-thought reasoning. But getting it configured properly took some trial and error. Here’s the complete setup guide.

Why Kimi For Coding?

Kimi K2.5 (via the Kimi For Coding API) offers:

• 262K context window — enough for large codebases
• 32K output tokens — can generate substantial code blocks
• Built-in reasoning — shows its work when solving problems
• Fast inference — noticeably quicker than some alternatives

The Configuration

Here’s my working clawdbot.json configuration:

{
  "env": {
    "vars": {
      "KIMICODE_API_KEY": "sk-kimi-YOUR_API_KEY_HERE"
    }
  },
  "models": {
    "mode": "merge",
    "providers": {
      "kimi-code": {
        "baseUrl": "https://api.kimi.com/coding/v1",
        "apiKey": "sk-kimi-YOUR_API_KEY_HERE",
        "api": "openai-completions",
        "models": [
          {
            "id": "kimi-for-coding",
            "name": "Kimi For Coding",
            "reasoning": true,
            "input": ["text"],
            "contextWindow": 262144,
            "maxTokens": 32768,
            "headers": {
              "User-Agent": "KimiCLI/0.77"
            },
            "compat": {
              "supportsDeveloperRole": false
            }
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "kimi-code/kimi-for-coding"
      },
      "thinkingDefault": "low",
      "models": {
        "kimi-code/kimi-for-coding": {
          "alias": "Kimi K2.5 (Kimi Code)"
        }
      }
    }
  }
}

Critical Configuration Details

1. The User-Agent Header (Required)

This is the gotcha that trips everyone up. Kimi restricts API access to approved coding agents. You must include a recognized User-Agent:

"headers": {
  "User-Agent": "KimiCLI/0.77"
}

Other working options:

• claude-code/1.0.0
• KimiCLI/0.77

Without this, you’ll get: 403 Kimi For Coding is currently only available for Coding Agents

2. Compatibility Flag

Kimi’s API doesn’t support the developer role that some OpenAI-compatible clients use. Set this to avoid role alternation errors:

"compat": {
  "supportsDeveloperRole": false
}

3. Reasoning Mode

Set "reasoning": true to enable chain-of-thought. The API returns two fields:

• content — the final answer you see
• reasoning_content — the model’s internal thinking process

4. Thinking Level

"thinkingDefault": "low"

Options: off, minimal, low, medium, high

Toggle in Telegram: /reasoning on or /reasoning off

Troubleshooting

”Message ordering conflict” Error

Cause: Clawdbot couldn’t parse the reasoning response format.

Fix: Ensure both are set:

• "reasoning": true in model config
• "compat": { "supportsDeveloperRole": false }

If it persists, clear sessions and restart:

# Backup first
tar -czvf ~/.clawdbot/sessions-backup.tar.gz -C ~/.clawdbot/agents/main sessions

# Clear and restart
rm -rf ~/.clawdbot/agents/main/sessions/*
npx clawdbot gateway stop && sleep 2 && npx clawdbot gateway start

403 Forbidden Error

Cause: Missing or incorrect User-Agent header.

Fix: Add the headers inside the model definition (not the provider level):

{
  "id": "kimi-for-coding",
  "headers": {
    "User-Agent": "KimiCLI/0.77"
  }
}

API Key Not Found

Note: The env.vars section does not feed into ${VAR} substitution for model configs. Put your API key directly in the provider config:

"apiKey": "sk-kimi-YOUR_ACTUAL_KEY"

Verify Everything Works

Via curl

curl -s -X POST "https://api.kimi.com/coding/v1/chat/completions" \
  -H "Authorization: Bearer sk-kimi-YOUR_KEY" \
  -H "User-Agent: KimiCLI/0.77" \
  -H "Content-Type: application/json" \
  -d '{"model":"kimi-for-coding","messages":[{"role":"user","content":"What is 47 * 83?"}]}' \
  | jq '.choices[0].message'

You should see both content and reasoning_content in the response.

Via logs

tail -100 /tmp/clawdbot/clawdbot-$(date +%Y-%m-%d).log | grep "thinking="

Look for: thinking=low (or your configured level)

Useful Commands

Summary

References

• Original gist with full config
• OpenCode Kimi Setup Gist
• Clawdbot GitHub

Got it working? The combination of local memory embeddings (from my previous post) plus Kimi’s reasoning makes for a surprisingly capable personal assistant setup.