Model Configuration

JeikCode decouples Provider Accounts (provider_accounts) from Model Profiles (models), supporting official APIs from major LLM vendors, proxy gateways, and local offline inference runtimes (such as Ollama, vLLM, and LM Studio).

---

1. Supported Provider Protocols

JeikCode natively supports the following upstream LLM communication protocols:

• openai-compatible: OpenAI Chat Completions compatible API and proxy gateways
• responses: OpenAI Responses native streaming protocol
• anthropic: Anthropic Claude native Messages protocol
• gemini: Google Gemini native protocol
• ollama: Local Ollama native protocol

---

2. Core TOML Model Configuration Template

Configure your provider accounts and models in ~/.jeikcode/config.toml:

# =============================================================================
# 1. Provider Accounts [provider_accounts.<account_id>]
# =============================================================================
[provider_accounts.deepseek]
provider = "deepseek"                       # Built-in presets: deepseek, openai, anthropic, zhipu, etc.
api_key = "sk-xxxxxxxxxxxxxxxxxxxxxxxx"
# base_url = "https://api.deepseek.com/v1" # Preset endpoints are built-in; override here if needed

[provider_accounts.custom-proxy]
provider = "openai-compatible"              # OpenAI Chat Completions protocol
api_key = "sk-xxxxxxxxxxxxxxxxxxxxxxxx"
base_url = "https://api.your-proxy.com/v1"

# =============================================================================
# 2. Model Profiles [models."<account_id>/<model_alias>"]
# =============================================================================
[models."deepseek-v4.1-flash"]
account = "deepseek"
model = "deepseek-v4.1-flash"              # Upstream actual model ID
context_window = 1024000                   # Context window token limit
max_tokens = 131200                        # Maximum completion tokens
reasoning_model = true                     # Declare as reasoning/thinking model
reasoning_history = "exclude"              # Thinking replay strategy: "include" | "exclude"
reasoning_effort = "high"                  # Default thinking budget: "low" | "medium" | "high" | "max"
reasoning_levels = ["low", "medium", "high", "max"] # Quick-cycle levels via Ctrl+T in terminal
image_input = true                         # Enable direct multimodal image input

---

3. Interactive Terminal Command Management

In the interactive terminal, you do not need to manually edit files every time; manage models and credentials via slash commands:

Switch Active Model (/model)

/model

Opens an interactive menu displaying configured models and your currently active primary model. Use arrow keys to quickly switch between models.

Configure Providers & API Keys (/provider)

/provider

Interactively add or edit provider accounts (DeepSeek, Anthropic Claude, OpenAI, Google Gemini, SiliconFlow, Zhipu AI, Moonshot, OpenRouter, Ollama, etc.) and enter your API key to write it to your configuration.

---

4. Environment Variables

You can also export API keys via environment variables. JeikCode reads them on startup:

Linux / macOS

# DeepSeek
export DEEPSEEK_API_KEY="sk-..."

# Anthropic Claude
export ANTHROPIC_API_KEY="sk-ant-..."

# OpenAI or Compatible Gateway
export OPENAI_API_KEY="sk-..."

# Google Gemini
export GEMINI_API_KEY="AIza..."

Windows (PowerShell)

# Current terminal session
$env:DEEPSEEK_API_KEY="sk-..."
$env:ANTHROPIC_API_KEY="sk-ant-..."
$env:OPENAI_API_KEY="sk-..."

# Permanent user environment variable
[Environment]::SetEnvironmentVariable("DEEPSEEK_API_KEY", "sk-...", "User")

---

5. Vision Preprocessor Fallback (vision_preprocessor_provider)

If your active primary model is a text-only coding model but you need to attach screenshots or design assets:

# Model profile used to preprocess and analyze image attachments
vision_preprocessor_provider = "custom/claude-sonnet"

The agent automatically routes attached images to this vision-capable model first, transcribing UI layouts and stack traces into structured descriptions before prompting the primary model.