Providers | Gokin

Gokin supports 7 AI model providers. All except Gemini and Ollama use an Anthropic-compatible API.

Overview#

Provider	Default Model	Auth	Context	Notes
Gemini	gemini-3-flash-preview	API Key + OAuth	1M	Free tier, Google OAuth
Anthropic	claude-sonnet-4-5-20250929	API Key	200K	Extended thinking
DeepSeek	deepseek-chat	API Key	64K+	Affordable reasoning
GLM	glm-4.7	API Key	128K	Coding Plan (Z.ai)
MiniMax	MiniMax-M2.5	API Key	1M	Huge context window
Kimi	kimi-k2.5	API Key	256K+	Thinking models
Ollama	llama3.2	Not required	Varies	Local models, offline

Gemini (Google)#

Models:

Model	Description	Cost
`gemini-3-flash-preview`	Fast and cheap	$0.50 / $3 per 1M tokens
`gemini-3-pro-preview`	Most powerful	$2 / $12 per 1M tokens
`gemini-2.5-flash`	Code Assist (OAuth)	Free
`gemini-2.5-pro`	Code Assist (OAuth)	Free

Environment variables (in priority order):

export GOKIN_GEMINI_KEY="your-key"
export GEMINI_API_KEY="your-key"
export GOOGLE_API_KEY="your-key"

Get a key: aistudio.google.com/apikey

OAuth for Gemini#

Gemini supports authorization via Google account without an API key:

# Login via OAuth
/oauth-login

# Logout
/oauth-logout

Uses the Code Assist API (cloudcode-pa.googleapis.com). Tokens are stored in configuration and refreshed automatically.

Anthropic (Claude)#

Models:

Model	Description
`claude-opus-4-6`	Most powerful Claude model
`claude-sonnet-4-5-20250929`	Balance of speed and quality
`claude-haiku-4-5-20251001`	Fast and affordable

Environment variables:

export GOKIN_ANTHROPIC_KEY="your-key"
export ANTHROPIC_API_KEY="your-key"

Get a key: console.anthropic.com/settings/keys

Extended Thinking:

model:
  enable_thinking: true
  thinking_budget: 4096

DeepSeek#

Models:

Model	Description
`deepseek-chat`	Powerful coding assistant
`deepseek-reasoner`	Extended reasoning

Environment variables:

export GOKIN_DEEPSEEK_KEY="your-key"
export DEEPSEEK_API_KEY="your-key"

Get a key: platform.deepseek.com/api_keys

GLM (Z.ai / Zhipu)#

Models:

Model	Description
`glm-5`	Latest GLM model
`glm-4.7`	Coding assistant, 131K max output

Environment variables:

export GOKIN_GLM_KEY="your-key"
export GLM_API_KEY="your-key"

Get a key: open.bigmodel.cn

MiniMax#

Models:

Model	Description
`MiniMax-M2.5`	Flagship: 1M context, strong coding

Environment variables:

export GOKIN_MINIMAX_KEY="your-key"
export MINIMAX_API_KEY="your-key"

Get a key: platform.minimaxi.com

Kimi (Moonshot AI)#

Models:

Model	Description
`kimi-k2.5`	Multimodal model
`kimi-k2-thinking-turbo`	Extended reasoning, 60-100 tok/s
`kimi-k2-turbo-preview`	Fast coding, 256K context

Environment variables:

export GOKIN_KIMI_KEY="your-key"
export KIMI_API_KEY="your-key"
export MOONSHOT_API_KEY="your-key"

Get a key: platform.moonshot.ai/console/api-keys

Ollama (Local Models)#

Ollama enables fully offline operation without API keys.

Requirements: running Ollama server on localhost:11434

Auto-detected model families:

Family	Example Models
Llama	`llama3.2`, `llama3.1`, `llama3`, `llama2`
Qwen	`qwen2.5-coder`, `qwen2.5`, `qwen2`
Mistral	`mistral-nemo`, `mistral`, `mixtral`
Phi	`phi4`, `phi3`
Code	`codellama`, `starcoder2`, `deepseek-coder`, `codegemma`
Other	`gemma2`, `command-r-plus`, `vicuna`, `yi`, `solar`

Environment variables (optional):

export GOKIN_OLLAMA_KEY="your-key"   # For remote server with auth
export OLLAMA_API_KEY="your-key"

Tool Call Fallback#

Models without native tool calling support use a text fallback: Gokin parses {"tool":"name","args":{}} JSON format from model responses.

Model Profiles#

Gokin automatically detects model capabilities:

Model	Context	Tools	Coding	Small
llama3.2	128K	Yes	No	Yes
qwen2.5-coder	32K	Yes	Yes	No
mistral-nemo	128K	Yes	No	No
phi4	16K	Yes	No	Yes
command-r-plus	128K	Yes	No	No

Switching Provider and Model#

# Via slash command
/provider anthropic
/model claude-sonnet-4-5-20250929

# Via configuration

# ~/.config/gokin/config.yaml
api:
  active_provider: gemini
model:
  name: gemini-3-flash-preview

Model Presets#

Preset	Description
`coding`	Optimized for code generation
`fast`	Minimum latency
`balanced`	Balance of quality and speed
`creative`	High temperature for generation

model:
  preset: coding

Provider Failover#

Gokin automatically switches to a fallback provider on errors:

model:
  name: claude-sonnet-4-5-20250929
  provider: anthropic
  fallback_providers:
    - deepseek
    - gemini
    - ollama

Provider health:

Each provider has a score from -20 to +8
Success: +1 to score
Fatal error: -2 to score
Data stored in ~/.config/gokin/provider_health.json
Providers with higher scores are tried first