Skip to main content
A lightweight AI agent CLI. Official repo: https://github.com/HKUDS/nanobot
Protocol: OpenAI-compatible
  • Base URL: https://api.orcarouter.ai/v1 (must include /v1)

​
Install

Requires uv first, then install the nanobot command: 
uv tool install nanobot-ai

​
Configure

Edit ~/.nanobot/config.json: 
{
  "agents": {
    "defaults": {
      "model": "orcarouter/auto",
      "provider": "custom",
      "maxTokens": 4096,
      "temperature": 0.1
    }
  },
  "providers": {
    "custom": {
      "apiKey": "sk-orca-...",
      "apiBase": "https://api.orcarouter.ai/v1"
    }
  }
}

​
Run

nanobot agent -m "Reply with exactly: OK" --no-logs --no-markdown
Expect the reply OK.

​
Notes

  • Use the predefined custom provider — do not invent your own provider name. Custom names are ignored and you get a provider 'None' error.
  • Do not add an apiType field to the custom provider. Only the built-in openai provider supports apiType; adding it to custom fails validation.
  • The base URL keeps the trailing /v1, and agents.defaults.provider must be set to custom.
  • Model names use vendor/model format. Swap orcarouter/auto for a specific model like anthropic/claude-opus-4.8 or google/gemini-2.5-flash.

​
Enable thinking / reasoning

Set reasoningEffort under agents.defaults (or a modelPresets entry). Valid values: none / low / medium / high / max. Omitting it (or null) follows the model’s default. 
{
  "agents": {
    "defaults": {
      "model": "anthropic/claude-sonnet-4.6",
      "provider": "custom",
      "maxTokens": 16000,
      "reasoningEffort": "high"
    }
  },
  "providers": { "custom": { "apiKey": "sk-orca-...", "apiBase": "https://api.orcarouter.ai/v1" } }
}
maxTokens must be larger than the thinking budget. nanobot translates reasoningEffort: high into an Anthropic thinking.budget_tokens; if maxTokens (e.g. the default 4096) is smaller, the upstream returns 400 max_tokens must be greater than thinking.budget_tokens. Bumping maxTokens to ~16000 fixes it. (Verified: returns ✻ 391.)