Providers

FreeChaOS is provider-agnostic. The kernel speaks the Chaos-ABI; adapters translate that to whatever wire format a given provider expects. New providers are added through ~/.chaos/config.toml - no code changes, no rebuilds.

Built-in providers

Provider	Wire format	Notes
openai	Responses API	Default, hardcoded
anthropic	Anthropic Messages	Hardcoded; URL-detected
xai	Responses API	Bundled; native web_search / x_search tools
zai	Chat Completions	Bundled; Z.ai pay-per-token (GLM-5 / GLM-5.1)
zai-coding	Chat Completions	Bundled; Z.ai GLM Coding Plan subscription
charm	Chat Completions	Bundled via thirdparty.toml

Any other provider - DeepSeek, Groq, Ollama, MiniMax, Kimi, TensorZero, self-hosted gateways - is a config entry away.

Wire formats

Providers speak one of four wire formats:

• Responses API - OpenAI’s /v1/responses. What OpenAI ships and most imitators clone.
• Chat Completions - /v1/chat/completions. The lingua franca of OpenAI-compatible gateways.
• Anthropic Messages - Anthropic’s native format. Auto-detected when the base URL contains anthropic.
• TensorZero - TensorZero’s native /inference endpoint. Opt in with wire_api = "tensorzero".

By default, providers use wire_api = "auto": FreeChaOS tries Responses first and falls back to Chat Completions on 404/405/501. The winning format is cached for the session.

Selection rules

1. If the base_url contains anthropic → Anthropic Messages API (overrides wire_api)
2. If wire_api is set explicitly → use it (responses, chat_completions, tensorzero)
3. Otherwise → auto: try Responses, fall back to Chat Completions on 404/405/501

There is no wire_api = "anthropic" option. The URL is the signal.

Examples

xAI (Grok)

Bundled - no config needed. Just export the key:

export XAI_API_KEY=xai-...
chaos --provider xai --model grok-4

URLs containing x.ai automatically expose xAI’s native web_search and x_search server-side tools - no function schema needed. Override the bundled config by redeclaring [model_providers.xai] in ~/.chaos/config.toml.

Anthropic (Claude)

Already built-in, but overridable:

[model_providers.anthropic]
name = "Anthropic"
base_url = "https://api.anthropic.com/v1"
env_key = "ANTHROPIC_API_KEY"

export ANTHROPIC_API_KEY=sk-ant-...
chaos --provider anthropic --model haiku

The URL contains anthropic, so FreeChaOS routes to the Messages API adapter.

Z.ai (GLM)

Bundled as two separate providers - Z.ai runs distinct endpoints for pay-per-token API access and the GLM Coding Plan subscription. Both share the same ZAI_API_KEY; pick the provider that matches your billing.

export ZAI_API_KEY=your-key
chaos --provider zai --model glm-5.1        # pay-per-token
chaos --provider zai-coding --model glm-5.1 # Coding Plan subscription

DeepSeek

[model_providers.deepseek]
name = "DeepSeek"
base_url = "https://api.deepseek.com/v1"
env_key = "DEEPSEEK_API_KEY"

Groq

[model_providers.groq]
name = "Groq"
base_url = "https://api.groq.com/openai/v1"
env_key = "GROQ_API_KEY"

Ollama (local)

[model_providers.ollama]
name = "Ollama"
base_url = "http://localhost:11434/v1"

No env_key needed - Ollama runs locally without authentication.

chaos --provider ollama --model llama3

TensorZero

[model_providers.tensorzero]
name = "TensorZero"
base_url = "http://localhost:3000"
wire_api = "tensorzero"

TensorZero has its own inference protocol - explicit wire_api required.

Anthropic-compatible proxies (MiniMax, Kimi, Z.ai)

Any provider whose base URL contains anthropic is routed to the Anthropic Messages adapter:

[model_providers.minimax]
name = "MiniMax"
base_url = "https://api.minimax.io/anthropic"
env_key = "MINIMAX_API_KEY"

Configuration reference

Field	Required	Description
name	yes	Display name for logs and model selection
base_url	yes	Provider API endpoint
env_key	no	Environment variable holding the API key
env_key_instructions	no	Help text shown when the key is missing
wire_api	no	"auto" (default), "responses", "chat_completions", or "tensorzero". Anthropic is URL-detected and overrides this.
http_headers	no	Static headers as { "Header-Name" = "value" }
env_http_headers	no	Headers from env vars as { "Header-Name" = "ENV_VAR" }
query_params	no	Query string parameters as { "key" = "value" }
request_max_retries	no	HTTP retry limit (default: 4, max: 100)
stream_max_retries	no	Stream reconnect limit (default: 5, max: 100)
stream_idle_timeout_ms	no	Idle timeout in ms (default: 300000)
supports_websockets	no	Enable WebSocket transport (default: false)
experimental_bearer_token	no	Hardcoded bearer token (discouraged - use env_key)

Troubleshooting

“env var not set” - Export the API key variable listed in env_key.

Provider returns errors - Check that base_url points to the correct API version endpoint. Most OpenAI-compatible providers use /v1.

Timeouts on slow providers - Increase stream_idle_timeout_ms:

[model_providers.slow]
stream_idle_timeout_ms = 600000

Rate limiting - FreeChaOS retries 429s automatically. Increase request_max_retries if needed.

Files

• ~/.chaos/config.toml - user-level provider configuration
• thirdparty.toml - bundled provider definitions referenced by the tree