orcx 0.0.5

LLM orchestrator - route prompts to any model via litellm

orcx

LLM orchestrator - route prompts to any model via litellm.

Alpha - Core functionality works, API may change.

Why orcx?

• 100+ providers via litellm - OpenRouter, OpenAI, Anthropic, Google, etc.
• Agent presets - Save model + system prompt + parameters as reusable configs
• Conversations - Multi-turn with -c continue and --resume ID
• OpenRouter routing - Control quantization, provider selection, sorting
• Cost tracking - See what each request costs
• Simple - Just route prompts, not a full agent framework

Install

uv tool install orcx
# or
pip install orcx

Quick Start

# Set your API key
export OPENROUTER_API_KEY=sk-or-...

# Run a prompt
orcx run -m openrouter/deepseek/deepseek-v3.2 "hello"

# With file context
orcx run -m openrouter/deepseek/deepseek-v3.2 -f code.py "review this"

# Show cost
orcx run -m openrouter/deepseek/deepseek-v3.2 "hello" --cost

Usage

# Direct model (provider/model format)
orcx run -m openrouter/deepseek/deepseek-v3.2 "hello"

# Using model alias (if configured)
orcx run -m deepseek "hello"

# Using agent preset
orcx run -a reviewer "review this code"

# With file context
orcx run -a reviewer -f src/main.py "review this"

# Multiple files
orcx run -m deepseek -f code.py -f tests.py "explain the tests"

# Pipe from stdin
cat code.py | orcx run -a reviewer "review this"

# JSON output (includes usage/cost)
orcx run -m deepseek "hello" --json

# Custom system prompt
orcx run -m deepseek -s "You are a pirate" "greet me"

# Save response to file
orcx run -m deepseek "explain async" -o response.md

Conversations

Conversations are saved automatically. Continue or resume them:

# Start a conversation (prints ID like [a1b2])
orcx run -m deepseek "explain python decorators"
# [a1b2]

# Continue last conversation
orcx run -c "show me an example"

# Resume specific conversation
orcx run --resume a1b2 "what about class decorators?"

# Don't save this exchange
orcx run --no-save -m deepseek "quick question"

# List recent conversations
orcx conversations

# Show full conversation
orcx conversations show a1b2

# Delete conversation
orcx conversations delete a1b2

# Clean old conversations (default: 30 days)
orcx conversations clean --days 7

Conversations are stored in ~/.config/orcx/conversations.db.

Configuration

Config location: ~/.config/orcx/

config.yaml

# Default model (used when no -m or -a specified)
default_model: openrouter/deepseek/deepseek-v3.2

# Model aliases for shorthand
aliases:
  deepseek: deepseek/deepseek-v3.2
  sonnet: anthropic/claude-4.5-sonnet

# Global provider preferences (OpenRouter only)
# Applied to all openrouter/* models, merged with agent-specific prefs
default_provider_prefs:
  min_bits: 8
  ignore: [SiliconFlow, DeepInfra]
  sort: price

# API keys (env vars take precedence)
keys:
  openrouter: sk-or-...

agents.yaml

agents:
  fast:
    model: openrouter/deepseek/deepseek-v3.2
    description: Fast, cheap reasoning
    temperature: 0.7
    max_tokens: 4096

  reviewer:
    model: anthropic/claude-4.5-sonnet
    system_prompt: You are a code reviewer. Be concise and actionable.

  # With OpenRouter provider preferences
  quality:
    model: openrouter/deepseek/deepseek-v3.2
    provider_prefs:
      min_bits: 8 # fp8+ only (excludes fp4)
      ignore: [DeepInfra] # blacklist providers
      prefer: [Together] # soft preference
      sort: throughput # or: price, latency

Provider Preferences

OpenRouter-specific routing options. Can be set globally in config.yaml (default_provider_prefs) or per-agent in agents.yaml (provider_prefs). Agent prefs are merged with global prefs (agent takes precedence).

Option	Type	Description
min_bits	int	Minimum quantization (8 = fp8+)
quantizations	list	Explicit whitelist: [fp8, fp16, bf16]
exclude_quants	list	Blacklist: [fp4, int4]
ignore	list	Blacklist providers
only	list	Whitelist providers (strict)
prefer	list	Soft preference (tries first, falls back)
order	list	Explicit order
sort	str	"price", "throughput", "latency"

Use --cost flag to see which prefs were applied to a request.

Commands

orcx run "prompt"        # Run prompt (uses default model/agent)
orcx run -m MODEL "..."  # Use specific model or alias
orcx run -a AGENT "..."  # Use agent preset
orcx run -c "..."        # Continue last conversation
orcx agents              # List configured agents
orcx models              # Show model format and examples
orcx conversations       # List/manage conversations
orcx --version           # Show version
orcx --debug             # Show full tracebacks on error

CLI Options

Option	Short	Description
--model	-m	Model or alias to use
--agent	-a	Agent preset to use
--system	-s	System prompt
--context		Context to prepend
--file	-f	Files to include (repeatable)
--output	-o	Write response to file
--continue	-c	Continue last conversation
--resume		Resume conversation by ID
--no-save		Don't save conversation
--no-stream		Disable streaming output
--cost		Show cost after response
--json	-j	Output as JSON

Environment Variables

API keys loaded from env vars (take precedence over config file):

• OPENROUTER_API_KEY
• OPENAI_API_KEY
• ANTHROPIC_API_KEY
• DEEPSEEK_API_KEY
• GOOGLE_API_KEY
• MISTRAL_API_KEY
• GROQ_API_KEY
• TOGETHER_API_KEY

Model Format

Models use provider/model-name format. See your provider's docs for available models:

• OpenRouter models
• litellm providers

License

MIT