omniagent

OmniAgent

One CLI to route AI tasks to the best model.

Stop switching between multiple AI CLIs manually. OmniAgent detects your task type, picks the optimal model for the job, and routes it automatically.

# Instead of using different tools for different tasks:
# tool-a "analyze this code"
# tool-b "translate this text"
# tool-c "write a function"

# Do this:
oa "review this code for security issues"   # → routes to DeepSeek (code review)
oa "translate this to Spanish"              # → routes to Groq (translation)
oa "write a component"                      # → routes to Groq (code generation)
oa "solve this equation"                    # → routes to DeepSeek (math)

Install

pip install -e .
# or
pipx install .

Requires Python 3.10+. The oa command is available after install.

Quick Start

# Set API keys for the providers you have
$env:GROQ_API_KEY="your-key"          # PowerShell
export GROQ_API_KEY="your-key"        # bash

# Or store keys persistently via the setup wizard
oa setup

# Run
oa "your task here"

# See routing decisions
oa run --verbose "your task here"

# Force a specific provider
oa run --provider groq "your task here"

# Preview routing without executing
oa run --dry-run "your task here"

# List available providers
oa providers-list

# Show cost breakdown
oa cost

Supported Providers

Provider	Model	Cost Tier	Best For
Groq	llama-3.3-70b-versatile	economy	Code gen, quick, budget
Google	gemini-2.0-flash	free	Research, general
OpenRouter	qwen-2.5-coder-32b-instruct	variable	Quality/cost balance
DeepSeek	deepseek-chat	economy	Code review, math, reasoning
Ollama	qwen2.5-coder:7b (local)	free-local	Offline, privacy

API keys are read from environment variables (GROQ_API_KEY, GOOGLE_API_KEY, OPENROUTER_API_KEY, DEEPSEEK_API_KEY) or from ~/.config/omniagent/config.yaml. Ollama needs no key — just a running local server at http://localhost:11434.

How Routing Works

OmniAgent classifies the prompt into a task type (code review, code generation, translation, math, quick query, creative writing, data analysis, research, general) using keyword signals. Each task type maps to a preferred provider; if that provider is unavailable, it falls back through an ordered list.

Task Type	Preferred Provider
Code review	deepseek
Code generation	groq
Translation	groq
Math	deepseek
Quick query	groq
Creative writing	openrouter
Data analysis	google
Research	google
General	groq

Fallback order: groq → google → deepseek → openrouter → ollama.

Cost Tracking

oa cost                # last 10 requests
oa cost --last 20      # custom window
oa cost --summary      # totals per provider
oa budget              # show monthly budget
oa budget --monthly 50 # set budget

Configuration

Stored at ~/.config/omniagent/config.yaml. Key options:

Option	Description
monthly_budget	Monthly spending cap (EUR)
routing.<task_type>	Preferred provider per task type
models.<provider>	Override the model per provider
ollama_base_url	Ollama server URL (default localhost:11434)

Tests

python -m pytest tests -q

Roadmap

• Parallel execution (split task across providers)
• Cost comparison mode (run same task on 3 providers, compare)
• Memory persistence across sessions
• MCP server integration
• Batch processing mode
• Web UI dashboard

License

MIT

This site is open source. Improve this page.