Skip to content

OpenAI-compatible provider

anthrogo supports any OpenAI Chat Completions-compatible endpoint via profile type: openai. This covers DeepSeek, Kimi, MiniMax, GLM, vllm, and many others.

Configuration via profiles

provider: deepseek           # set active profile
profiles:
  deepseek:
    type: openai
    base_url: https://api.deepseek.com
    model: deepseek-chat
    api_key: env:DEEPSEEK_API_KEY
  kimi:
    type: openai
    base_url: https://api.moonshot.cn/v1
    model: kimi-k2-0905-preview
    api_key: env:KIMI_API_KEY
  minimax:
    type: openai
    base_url: https://api.minimaxi.com/v1
    model: MiniMax-M2
    api_key: env:MINIMAX_API_KEY
  glm:
    type: openai
    base_url: https://open.bigmodel.cn/api/paas/v4
    model: glm-4.6
    api_key: env:GLM_API_KEY

Known providers tested

Provider type base_url Model examples
DeepSeek openai https://api.deepseek.com deepseek-chat, deepseek-reasoner
Kimi openai https://api.moonshot.cn/v1 kimi-k2-0905-preview
MiniMax (OpenAI-compat) openai https://api.minimaxi.com/v1 MiniMax-M2, abab6.5s-chat
MiniMax (Anthropic-compat) anthropic https://api.minimaxi.com/anthropic minimax-m2.7 (trailing slash auto-normalized)
GLM openai https://open.bigmodel.cn/api/paas/v4 glm-4.6, glm-zero-preview
Google Gemini openai https://generativelanguage.googleapis.com/v1beta/openai gemini-2.5-flash, gemini-2.5-flash-lite, gemini-flash-latest, gemini-3-flash-preview
vllm openai http://localhost:8000/v1 any loaded model
Ollama ollama n/a (use type: ollama) local models — prefer the dedicated ollama provider

Notes for Gemini via OpenAI-compat

  • api_key: a Google AI Studio key (AIza…). Set api_key: env:GEMINI_API_KEY in the profile.
  • Free tier limits: ~20 requests/day/model on gemini-2.5-flash and gemini-2.5-flash-lite. Each model has independent quota — anthrogo will switch + retry automatically with the v0.14.2 retry-after-aware backoff.
  • Gemini 3 / gemini-flash-latest: requires thought_signature echo on multi-turn tool use. anthrogo passes this through via ProviderMetadata on tool_use blocks (since v0.14.2). Without v0.14.2+, second-turn tool calls fail with HTTP 400.
  • reasoning_content from Gemini 3's thinking is surfaced as EventThinkingDelta (separate from assistant text). TUI renders it dimmed; headless --json mode emits {"kind":"thinking_delta", ...}.

Switching profiles at runtime

anthrogo --provider kimi
anthrogo --provider deepseek -p "summarize this repo"

Built-in pricing defaults

Model Input ($/M) Output ($/M)
deepseek-chat 0.27 1.10
deepseek-reasoner — —
kimi-k2* — —
MiniMax-M2 — —
glm-4 / glm-zero- — —

Add explicit pricing: entries for unlisted models or negotiated rates.

Vision support

OpenAI-compatible image blocks are converted to {type: "image_url", image_url: {url: "data:<mime>;base64,..."}} multimodal content arrays. Use @image:<path> in your prompt.