Skip to main content
The routing configuration determines which LLM model each process type uses. You can set defaults in [defaults.routing] and override per agent in [agents.routing].

​
Process-Type Routing

​
channel
string
default:"anthropic/claude-sonnet-4"
Model used by channel processes (user-facing conversations).
[defaults.routing]
channel = "anthropic/claude-sonnet-4"
​
branch
string
default:"anthropic/claude-sonnet-4"
Model used by branch processes (forked thinking).
[defaults.routing]
branch = "anthropic/claude-sonnet-4"
​
worker
string
default:"anthropic/claude-sonnet-4"
Model used by worker processes (task execution).
[defaults.routing]
worker = "anthropic/claude-haiku-4.5"
​
compactor
string
default:"anthropic/claude-sonnet-4"
Model used by compactor processes (context summarization).
[defaults.routing]
compactor = "anthropic/claude-haiku-4.5"
​
cortex
string
default:"anthropic/claude-sonnet-4"
Model used by cortex processes (system observation, bulletin generation).
[defaults.routing]
cortex = "anthropic/claude-haiku-4.5"
​
voice
string
Model used for voice processing (if enabled).
[defaults.routing]
voice = "openai/gpt-4.1-mini"

​
Task-Type Overrides

​
task_overrides
object
Task-specific model overrides for workers and branches. When a worker or branch is spawned with a specific task type, use the override model instead of the default.
[defaults.routing.task_overrides]
coding = "anthropic/claude-sonnet-4"
research = "openai/gpt-4.1"
memory = "anthropic/claude-haiku-4.5"

​
Thinking Effort

​
channel_thinking_effort
string
default:"auto"
Thinking effort level for channel models. Controls extended thinking tokens.Valid values: auto, low, medium, high
[defaults.routing]
channel_thinking_effort = "medium"
​
branch_thinking_effort
string
default:"auto"
Thinking effort level for branch models.
[defaults.routing]
branch_thinking_effort = "high"
​
worker_thinking_effort
string
default:"auto"
Thinking effort level for worker models.
[defaults.routing]
worker_thinking_effort = "medium"
​
compactor_thinking_effort
string
default:"auto"
Thinking effort level for compactor models.
[defaults.routing]
compactor_thinking_effort = "low"
​
cortex_thinking_effort
string
default:"auto"
Thinking effort level for cortex models.
[defaults.routing]
cortex_thinking_effort = "medium"

​
Fallback Chains

​
fallbacks
object
Fallback chains for resilience. When a model fails with a retriable error (429, 502, 503, 504, rate limit), try the next model in its chain.
[defaults.routing.fallbacks]
"anthropic/claude-sonnet-4" = ["anthropic/claude-haiku-4.5", "openai/gpt-4.1"]
"openai/gpt-4.1" = ["openai/gpt-4.1-mini"]
​
rate_limit_cooldown_secs
integer
default:"60"
How long to deprioritize a rate-limited model (seconds). After a rate limit error, the model is avoided for this duration.
[defaults.routing]
rate_limit_cooldown_secs = 120

​
Model Name Format

Model names follow the format provider/model-id. The provider must match a configured LLM provider or custom provider.

​
Built-in Providers

  • anthropic/ - Anthropic models (e.g., anthropic/claude-sonnet-4)
  • openai/ - OpenAI models (e.g., openai/gpt-4.1)
  • openrouter/ - OpenRouter models (e.g., openrouter/anthropic/claude-sonnet-4-20250514)
  • kilo/ - Kilo Gateway models (e.g., kilo/anthropic/claude-sonnet-4.5)
  • gemini/ - Google Gemini models (e.g., gemini/gemini-2.0-flash-exp)
  • groq/ - Groq models (e.g., groq/llama-3.3-70b-versatile)
  • deepseek/ - DeepSeek models (e.g., deepseek/deepseek-chat)
  • xai/ - xAI models (e.g., xai/grok-2-latest)
  • mistral/ - Mistral AI models (e.g., mistral/mistral-large-latest)
  • together/ - Together AI models
  • fireworks/ - Fireworks AI models
  • nvidia/ - NVIDIA NIM models
  • ollama/ - Ollama models (e.g., ollama/llama3.3)

​
Custom Providers

For custom providers defined in [llm.provider.<name>], use <name>/model-id: 
[llm.provider.my_proxy]
api_type = "openai_completions"
base_url = "https://my-proxy.example.com/v1"
api_key = "env:MY_PROXY_KEY"

[defaults.routing]
channel = "my_proxy/gpt-4"

​
Examples

​
Single Provider (Anthropic)

[llm]
anthropic_key = "env:ANTHROPIC_API_KEY"

[defaults.routing]
channel = "anthropic/claude-sonnet-4"
branch = "anthropic/claude-sonnet-4"
worker = "anthropic/claude-haiku-4.5"
compactor = "anthropic/claude-haiku-4.5"
cortex = "anthropic/claude-haiku-4.5"

​
Mixed Providers with Task Overrides

[llm]
anthropic_key = "env:ANTHROPIC_API_KEY"
openai_key = "env:OPENAI_API_KEY"

[defaults.routing]
channel = "anthropic/claude-sonnet-4"
branch = "anthropic/claude-sonnet-4"
worker = "anthropic/claude-haiku-4.5"
compactor = "openai/gpt-4.1-mini"
cortex = "openai/gpt-4.1-mini"

[defaults.routing.task_overrides]
coding = "anthropic/claude-sonnet-4"
research = "openai/gpt-4.1"

​
OpenRouter with Fallbacks

[llm]
openrouter_key = "env:OPENROUTER_API_KEY"

[defaults.routing]
channel = "openrouter/anthropic/claude-sonnet-4-20250514"
branch = "openrouter/anthropic/claude-sonnet-4-20250514"
worker = "openrouter/anthropic/claude-haiku-4.5-20250514"
compactor = "openrouter/anthropic/claude-haiku-4.5-20250514"
cortex = "openrouter/anthropic/claude-haiku-4.5-20250514"

[defaults.routing.fallbacks]
"openrouter/anthropic/claude-sonnet-4-20250514" = [
  "openrouter/anthropic/claude-haiku-4.5-20250514",
  "openrouter/google/gemini-2.0-flash-exp"
]

​
Per-Agent Routing Override

[defaults.routing]
channel = "anthropic/claude-haiku-4.5"
worker = "anthropic/claude-haiku-4.5"

# Default agent uses fast, cheap models
[[agents]]
id = "support"
default = true

# Coding agent uses more powerful models
[[agents]]
id = "coding"

[agents.routing]
channel = "anthropic/claude-sonnet-4"
worker = "anthropic/claude-sonnet-4"

[agents.routing.task_overrides]
coding = "anthropic/claude-sonnet-4"

​
Thinking Effort Configuration

[defaults.routing]
channel = "anthropic/claude-sonnet-4"
branch = "anthropic/claude-sonnet-4"
worker = "anthropic/claude-sonnet-4"

# Use extended thinking for complex tasks
channel_thinking_effort = "medium"
branch_thinking_effort = "high"
worker_thinking_effort = "medium"

# Keep compaction and cortex lightweight
compactor_thinking_effort = "low"
cortex_thinking_effort = "low"

​
Retriable Errors

The following errors trigger fallback to the next model in the chain:
  • HTTP 429 (rate limit)
  • HTTP 502 (bad gateway)
  • HTTP 503 (service unavailable)
  • HTTP 504 (gateway timeout)
  • Connection errors
  • Timeouts
  • Empty responses
  • Malformed responses
Context overflow errors (token limit exceeded) do NOT trigger fallbacks - they trigger compaction instead.

Build docs developers (and LLMs) love

Get started for freeTalk to us