NAME
ROUTING
SYNOPSIS
/route <add NAME URL [MODEL]|rm NAME|models URL|probe NAME|set NAME tools on|off|list|on|off|status>
@name prompt text
DESCRIPTION
Harvey can dispatch individual prompts to remote LLM endpoints — other Ollama instances on a Pi cluster, Llamafile servers, or cloud providers — using @mention syntax. Prefix any prompt with @name to send it to the named endpoint instead of the local model. The reply is streamed back and lands in the local conversation history so future turns retain full context.
Routing is explicitly user-driven: there is no automatic classification. You choose which endpoint handles each prompt by using (or omitting) an @mention.
CONTEXT WINDOW
When a prompt is dispatched to a remote endpoint, Harvey sends the last 10 non-system messages from the local history alongside it. System messages are excluded. This gives the remote model enough context to be useful without sending the entire conversation over the network. The window size is a starting point and will be tuned over time.
ENDPOINT TYPES
Local providers (no API key):
ollama://host:port A remote Ollama server (also accepts http:// and https://). llamafile://host:port A Llamafile binary server (OpenAI-compatible, port 8080). llamacpp://host:port A llama.cpp server (OpenAI-compatible, port 8080).
Cloud providers (API key read from environment):
anthropic:// Anthropic Claude (ANTHROPIC_API_KEY) deepseek:// DeepSeek (DEEPSEEK_API_KEY) gemini:// Google Gemini (GEMINI_API_KEY or GOOGLE_API_KEY) mistral:// Mistral (MISTRAL_API_KEY) openai:// OpenAI (OPENAI_API_KEY)
EXAMPLE SESSION
# Register a Pi cluster node
/route add pi2 ollama://192.168.1.12:11434 llama3.1:8b
# Register the Anthropic cloud endpoint
/route add claude anthropic:// claude-sonnet-4-20250514
# Enable routing
/route on
# Dispatch a complex task to the cloud
@claude refactor this module to use the repository pattern
# Run a quick task on a Pi node
@pi2 write a unit test for the Parse function
# Local model handles everything else (no @mention)
what does this error mean?
SLASH COMMANDS
/route add NAME URL [MODEL] register a remote endpoint
@pi2 ollama://192.168.1.12:11434 llama3.1:8b
@claude anthropic:// claude-sonnet-4-20250514
/route remove NAME unregister an endpoint (alias: rm)
/route rm NAME alias for remove
/route use [NAME] set NAME as the sticky route so all prompts
dispatch there without an @mention; omit NAME
to clear the sticky route
/route models URL list models available at a provider URL
useful before /route add to choose a model
/route probe NAME show reachability, model, and tool-call capability
for a registered endpoint
/route set NAME tools on|off toggle tool calling for a registered endpoint
(only for providers that support tool use)
/route list show all endpoints with reachability status
/route on enable @mention dispatch (persisted)
/route off disable @mention dispatch (persisted)
/route status show routing state and endpoint count
Registered endpoints and the on/off state persist across sessions in
<workspace>/agents/routes.json.
SEE ALSO
/model use NAME — backend-agnostic model switching /llamafile — local llamafile backend management /help security — API key filtering for cloud endpoints