llm-orchestra 0.15.11

Multi-agent LLM communication system with ensemble orchestration

LLM Orchestra

Orchestrate ensembles of specialized models — local and cloud — to do real analytical work. Coordination over scale.

Overview

A decent laptop can run multiple small language models simultaneously. What's missing is the coordination layer — the system that decomposes problems, routes them to specialized agents, manages dependencies between them, and synthesizes results. LLM Orchestra provides that layer.

The approach is architectural intelligence over brute-force scaling: instead of sending everything to one large model, decompose the problem and let specialized agents own their piece. Independent agents run in parallel. Dependent agents wait for what they need. Script agents handle data processing and analysis alongside LLM agents, enabling hybrid workflows that go beyond pure language model orchestration.

Mix expensive cloud models with free local models. Use Claude for strategic synthesis while local models handle systematic analysis at zero marginal cost.

Key Features

• Multi-Agent Ensembles: Coordinate specialized agents with flexible dependency graphs
• Ensemble Agents: Compose ensembles hierarchically — agents can reference and execute other ensembles
• Input Key Routing: Select specific keys from upstream JSON output for classify → route → fan-out patterns
• Agent Dependencies: Define which agents depend on others for sophisticated orchestration patterns
• Script Agent Integration: Execute custom scripts alongside LLM agents with JSON I/O communication
• Model Profiles: Simplified configuration with named shortcuts for model + provider combinations
• Cost Optimization: Mix expensive and free models based on what each task needs
• Streaming Output: Real-time progress updates during ensemble execution
• CLI Interface: Simple commands with piping support (cat code.py | llm-orc invoke code-review)
• Secure Authentication: Encrypted API key storage with easy credential management
• YAML Configuration: Easy ensemble setup with readable config files
• Usage Tracking: Token counting, cost estimation, and timing metrics
• Artifact Management: Automatic saving of execution results with timestamped persistence

Installation

Option 1: Homebrew (macOS - Recommended)

# Add the tap
brew tap mrilikecoding/llm-orchestra

# Install LLM Orchestra
brew install llm-orchestra

# Verify installation
llm-orc --version

Option 2: pip (All Platforms)

# Install from PyPI
pip install llm-orchestra

# Verify installation
llm-orc --version

Option 3: Development Installation

# Clone the repository
git clone https://github.com/mrilikecoding/llm-orc.git
cd llm-orc

# Install with development dependencies
uv sync --dev

# Verify installation
uv run llm-orc --version

Updates

# Homebrew users
brew update && brew upgrade llm-orchestra

# pip users
pip install --upgrade llm-orchestra

Quick Start

1. Set Up Authentication

Before using LLM Orchestra, configure authentication for your LLM providers:

# Interactive setup wizard (recommended for first-time users)
llm-orc auth setup

# Or add providers individually
llm-orc auth add anthropic --api-key YOUR_ANTHROPIC_KEY
llm-orc auth add google --api-key YOUR_GOOGLE_KEY

# OAuth for Claude Pro/Max users
llm-orc auth add anthropic-claude-pro-max

# List configured providers
llm-orc auth list

# Remove a provider if needed
llm-orc auth remove anthropic

Security: API keys are encrypted and stored securely in ~/.config/llm-orc/credentials.yaml.

2. Configuration Options

LLM Orchestra supports both global and local configurations:

Global Configuration

Create ~/.config/llm-orc/ensembles/code-review.yaml:

name: code-review
description: Multi-perspective code review ensemble

agents:
  - name: security-reviewer
    model_profile: free-local
    system_prompt: "You are a security analyst. Focus on identifying security vulnerabilities, authentication issues, and potential attack vectors."

  - name: performance-reviewer
    model_profile: free-local
    system_prompt: "You are a performance analyst. Focus on identifying bottlenecks, inefficient algorithms, and scalability issues."

  - name: quality-reviewer
    model_profile: free-local
    system_prompt: "You are a code quality analyst. Focus on maintainability, readability, and best practices."

  - name: senior-reviewer
    model_profile: default-claude
    depends_on: [security-reviewer, performance-reviewer, quality-reviewer]
    system_prompt: |
      You are a senior engineering lead. Synthesize the security, performance,
      and quality analysis into actionable recommendations.
    output_format: json

Local Project Configuration

For project-specific ensembles, initialize local configuration:

# Initialize local configuration in your project
llm-orc config init

# This creates .llm-orc/ directory with:
# - ensembles/   (project-specific ensembles)
# - models/      (shared model configurations)
# - scripts/     (project-specific scripts)
# - config.yaml  (project configuration)

View Current Configuration

# Check configuration status with visual indicators
llm-orc config check

3. Using LLM Orchestra

Basic Usage

# List available ensembles
llm-orc list-ensembles

# List available model profiles
llm-orc list-profiles

# Get help for any command
llm-orc --help
llm-orc invoke --help

Invoke Ensembles

# Analyze code from a file (pipe input)
cat mycode.py | llm-orc invoke code-review

# Provide input directly
llm-orc invoke code-review --input "Review this function: def add(a, b): return a + b"

# JSON output for integration with other tools
llm-orc invoke code-review --input "..." --output-format json

# Use specific configuration directory
llm-orc invoke code-review --config-dir ./custom-config

# Enable streaming for real-time progress (enabled by default)
llm-orc invoke code-review --streaming

Output Formats

LLM Orchestra supports three output formats for different use cases:

Rich Interface (Default)

Interactive format with real-time progress updates and visual dependency graphs:

llm-orc invoke code-review --input "def add(a, b): return a + b"

JSON Output

Structured data format for integration and automation:

llm-orc invoke code-review --output-format json --input "code to review"

Returns complete execution data including events, results, metadata, and dependency information.

Text Output

Clean, pipe-friendly format for command-line workflows:

llm-orc invoke code-review --output-format text --input "code to review"

Plain text results perfect for piping and scripting: llm-orc invoke ... | grep "security"

Configuration Management

# Initialize local project configuration
llm-orc config init --project-name my-project

# Check configuration status with visual indicators
llm-orc config check                # Global + local status with legend
llm-orc config check-global        # Global configuration only  
llm-orc config check-local         # Local project configuration only

# Reset configurations with safety options
llm-orc config reset-global        # Reset global config (backup + preserve auth by default)
llm-orc config reset-local         # Reset local config (backup + preserve ensembles by default)

# Advanced reset options
llm-orc config reset-global --no-backup --reset-auth       # Complete reset including auth
llm-orc config reset-local --reset-ensembles --no-backup   # Reset including ensembles

Script Management

LLM Orchestra includes powerful script agent integration for executing custom scripts alongside LLM agents:

# List available scripts in your project
llm-orc scripts list

# Show detailed information about a script
llm-orc scripts show file_operations/read_file.py

# Test a script with parameters
llm-orc scripts test file_operations/read_file.py --parameters '{"filepath": "example.txt"}'

# Scripts are discovered from .llm-orc/scripts/ directories
# Results are automatically saved to .llm-orc/artifacts/ with timestamps

Script agents use JSON I/O for seamless integration with LLM agents, enabling powerful hybrid workflows where scripts provide data and context for LLM analysis.

MCP Server

LLM Orchestra includes a Model Context Protocol (MCP) server that exposes ensembles, artifacts, and metrics as MCP resources. This enables integration with MCP clients like Claude Code, Claude Desktop, and other tools.

Quick Start

1. Add .mcp.json to your project root:

{
  "mcpServers": {
    "llm-orc": {
      "command": "uv",
      "args": ["run", "llm-orc", "mcp", "serve"]
    }
  }
}

2. Restart Claude Code - MCP tools appear as mcp__llm-orc__*

3. Try it:

mcp__llm-orc__get_help              # Get full documentation
mcp__llm-orc__get_provider_status   # Check which models are available
mcp__llm-orc__list_ensembles        # See available ensembles

Resources (Read-Only Data)

Resource	Description
llm-orc://ensembles	List all available ensembles with metadata
llm-orc://ensemble/{name}	Get specific ensemble configuration
llm-orc://profiles	List model profiles
llm-orc://artifacts/{ensemble}	List execution artifacts for an ensemble
llm-orc://artifact/{ensemble}/{id}	Get individual artifact details
llm-orc://metrics/{ensemble}	Get aggregated metrics (success rate, cost, duration)

Tools (25 Total)

Core Execution

Tool	Description
invoke	Execute ensemble with streaming progress, saves artifacts automatically
list_ensembles	List all ensembles from local/library/global sources
validate_ensemble	Check config validity, profile availability, and dependencies
update_ensemble	Modify ensemble config (supports dry-run and backup)
analyze_execution	Analyze execution artifact data

Provider Discovery - Check what's available before running

Tool	Description
get_provider_status	Show available providers and Ollama models
check_ensemble_runnable	Check if ensemble can run, suggest local alternatives

Ensemble Management

Tool	Description
create_ensemble	Create new ensemble from scratch or template
delete_ensemble	Delete ensemble (requires confirmation)

Profile Management

Tool	Description
list_profiles	List profiles with optional provider filter
create_profile	Create new model profile
update_profile	Update existing profile
delete_profile	Delete profile (requires confirmation)

Script Management

Tool	Description
list_scripts	List primitive scripts by category
get_script	Get script source and metadata
test_script	Test script with sample input
create_script	Create new primitive script
delete_script	Delete script (requires confirmation)

Library Operations

Tool	Description
library_browse	Browse library ensembles and scripts
library_copy	Copy from library to local project
library_search	Search library by keyword
library_info	Get library metadata and statistics

Library tools require a local copy of the library. These tools read from the local filesystem — they do not fetch from GitHub. The library is auto-detected if the llm-orchestra-library submodule is present in the current working directory. For Homebrew or pip installs, set LLM_ORC_LIBRARY_PATH=/path/to/llm-orchestra-library to point to a local clone of llm-orchestra-library. Run library_info to verify the library is found.

Artifact Management

Tool	Description
delete_artifact	Delete individual execution artifact
cleanup_artifacts	Delete old artifacts (supports dry-run)

Help

Tool	Description
get_help	Get comprehensive docs: directory structure, schemas, workflows

Example Workflow

# 1. Check what's available
mcp__llm-orc__get_provider_status
# → Shows Ollama running with llama3, mistral models

# 2. Find an ensemble
mcp__llm-orc__library_search query="code review"
# → Returns results including path "ensembles/code-analysis/security-review.yaml"

# 3. Check if it can run locally
mcp__llm-orc__check_ensemble_runnable ensemble_name="security-review"
# → Shows which profiles need local alternatives

# 4. Copy and adapt (source path is relative to library root)
mcp__llm-orc__library_copy source="ensembles/code-analysis/security-review"
mcp__llm-orc__update_ensemble ensemble_name="security-review" changes={"agents": [...]}

# 5. Run it
mcp__llm-orc__invoke ensemble_name="security-review" input_data="Review this code..."

CLI Usage

# Start MCP server (stdio transport for MCP clients)
llm-orc mcp serve

# Start with HTTP transport for debugging
llm-orc mcp serve --transport http --port 8080

Ensemble Library

Looking for pre-built ensembles? Check out the LLM Orchestra Library - a curated collection of analytical ensembles for code review, research analysis, decision support, and more.

Library CLI Commands

LLM Orchestra includes built-in commands to browse and copy ensembles from the library:

# Browse all available categories
llm-orc library categories
llm-orc l categories  # Using alias

# Browse ensembles in a specific category
llm-orc library browse code-analysis

# Show detailed information about an ensemble
llm-orc library show code-analysis/security-review

# Copy an ensemble to your local configuration
llm-orc library copy code-analysis/security-review

# Copy an ensemble to your global configuration
llm-orc library copy code-analysis/security-review --global

Library Source Configuration

By default, LLM Orchestra uses local filesystem detection: it checks for a llm-orchestra-library/ directory in the current working directory, then falls back to a no-op if none is found. Remote GitHub is not used unless explicitly configured. To use a specific library source:

# Use a local library at a custom path
export LLM_ORC_LIBRARY_PATH=/path/to/llm-orchestra-library
llm-orc library browse research-analysis

# Use local package submodule explicitly
export LLM_ORC_LIBRARY_SOURCE=local
llm-orc library browse research-analysis  # Uses local submodule
llm-orc init                              # Copies from local submodule

# Use remote GitHub library
export LLM_ORC_LIBRARY_SOURCE=remote
llm-orc library browse research-analysis

When to use local library:

• Testing changes to library ensembles before publishing
• Working on feature branches of the llm-orchestra-library
• Offline development (when remote access unavailable)
• Custom ensemble development and testing

Requirements for local library:

• The llm-orchestra-library submodule must be initialized and present
• Clear error messages guide you if the local library is not found

Contributing to the Library

The library is a separate repository at github.com/mrilikecoding/llm-orchestra-library. To add or improve content:

1. Create and test your ensemble locally using llm-orc or the MCP tools
2. Copy the finished YAML into your local library clone under the appropriate category
3. Open a pull request against the library repository

The MCP library_copy tool copies from the library to your project. There is no reverse direction by design — contributing back goes through a PR rather than automated writes to a shared upstream.

Use Cases

Code Review

Get systematic analysis across security, performance, and maintainability dimensions. Each agent focuses on their specialty while synthesis provides actionable recommendations.

Architecture Review

Analyze system designs from scalability, security, performance, and reliability perspectives. Identify bottlenecks and suggest architectural patterns.

Product Strategy

Evaluate business decisions from market, financial, competitive, and user experience angles. Get comprehensive analysis for complex strategic choices.

Research Analysis

Systematic literature review, methodology evaluation, or multi-dimensional analysis of research questions.

Model Support

• Claude (Anthropic) - Strategic analysis and synthesis
• Gemini (Google) - Multi-modal and reasoning tasks
• Ollama - Local deployment of open-source models (Llama3, etc.)
• Custom models - Extensible interface for additional providers

Configuration

Model Profiles

Model profiles simplify ensemble configuration by providing named shortcuts for complete agent configurations including model, provider, system prompts, timeouts, and generation parameters:

# In ~/.config/llm-orc/config.yaml or .llm-orc/config.yaml
model_profiles:
  free-local:
    model: llama3
    provider: ollama
    cost_per_token: 0.0
    system_prompt: "You are a helpful assistant that provides concise, accurate responses for local development and testing."
    timeout_seconds: 30
    temperature: 0.7
    max_tokens: 500
    options:              # Provider-specific parameters (Ollama)
      num_ctx: 4096
      top_k: 40

  default-claude:
    model: claude-sonnet-4-20250514
    provider: anthropic-claude-pro-max
    system_prompt: "You are an expert assistant that provides high-quality, detailed analysis and solutions."
    timeout_seconds: 60
    temperature: 0.5
    max_tokens: 2000

  high-context:
    model: claude-3-5-sonnet-20241022
    provider: anthropic-api
    cost_per_token: 3.0e-06
    system_prompt: "You are an expert assistant capable of handling complex, multi-faceted problems with detailed analysis."
    timeout_seconds: 120

  small:
    model: claude-3-haiku-20240307
    provider: anthropic-api
    cost_per_token: 1.0e-06
    system_prompt: "You are a quick, efficient assistant that provides concise and accurate responses."
    timeout_seconds: 30

Profile Benefits:

• Complete Agent Configuration: Includes model, provider, system prompts, timeout settings, and generation parameters
• Simplified Configuration: Use model_profile: default-claude instead of explicit model + provider + system_prompt + timeout
• Consistency: Same profile names work across all ensembles with consistent behavior
• Cost Tracking: Built-in cost information for budgeting
• Generation Control: Set temperature, max_tokens, and provider-specific options per profile
• Flexibility: Local profiles override global ones, explicit agent configs override profile defaults

Usage in Ensembles:

agents:
  - name: bulk-analyzer
    model_profile: free-local     # Complete config: model, provider, prompt, timeout
  - name: expert-reviewer
    model_profile: default-claude # High-quality config with appropriate timeout
  - name: document-processor
    model_profile: high-context   # Large context processing with extended timeout
    system_prompt: "Custom prompt override"  # Overrides profile default

Override Behavior: Explicit agent configuration takes precedence over model profile defaults:

agents:
  - name: custom-agent
    model_profile: free-local
    system_prompt: "Custom prompt"  # Overrides profile system_prompt
    timeout_seconds: 60            # Overrides profile timeout_seconds
    temperature: 0.1               # Overrides profile temperature
    max_tokens: 200                # Overrides profile max_tokens
    options:                       # Merged with profile options (agent wins)
      num_ctx: 8192

Ensemble Configuration

Ensemble configurations support:

• Model profiles for simplified, consistent model selection
• Agent specialization with role-specific prompts
• Generation parameters (temperature, max_tokens, options) per profile or per agent
• Agent dependencies using depends_on for sophisticated orchestration
• Dependency validation with automatic cycle detection and missing dependency checks
• Timeout management per agent with performance configuration
• Mixed model strategies combining local and cloud models
• Output formatting (text, JSON) for integration
• Streaming execution with real-time progress updates

Agent Dependencies

The new dependency-based architecture allows agents to depend on other agents, enabling sophisticated orchestration patterns:

agents:
  # Independent agents execute in parallel
  - name: security-reviewer
    model_profile: free-local
    system_prompt: "Focus on security vulnerabilities..."

  - name: performance-reviewer  
    model_profile: free-local
    system_prompt: "Focus on performance issues..."

  # Dependent agent waits for dependencies to complete
  - name: senior-reviewer
    model_profile: default-claude
    depends_on: [security-reviewer, performance-reviewer]
    system_prompt: "Synthesize the security and performance analysis..."

Benefits:

• Flexible orchestration: Create complex dependency graphs beyond simple coordinator patterns
• Parallel execution: Independent agents run concurrently for better performance
• Automatic validation: Circular dependencies and missing dependencies are detected at load time
• Better maintainability: Clear, explicit dependencies instead of implicit coordinator relationships

Fan-Out (Parallel Map-Reduce)

Agents with fan_out: true automatically expand into N parallel instances when their upstream dependency produces an array result. This enables map-reduce style parallel processing:

agents:
  # "Map" step: split input into chunks
  - name: chunker
    script: scripts/chunker.py
    # Returns: {"success": true, "data": ["chunk1", "chunk2", "chunk3"]}

  # "Reduce" step: process each chunk in parallel
  - name: processor
    model_profile: default-local
    depends_on: [chunker]
    fan_out: true
    system_prompt: "Analyze this text chunk..."

  # Synthesis: combine all results
  - name: synthesizer
    model_profile: default-local
    depends_on: [processor]
    system_prompt: "Synthesize the analysis results..."

How it works:

1. chunker runs and returns a JSON array (direct array or {"data": [...]} format)
2. processor is expanded into processor[0], processor[1], processor[2] — one per array element
3. All instances execute in parallel, each receiving their chunk plus metadata (chunk_index, total_chunks, base_input)
4. Results are gathered back under the original processor name as an ordered array
5. synthesizer receives the combined results and can reference them normally

Configuration requirements:

• fan_out: true requires a depends_on field (validated at load time)
• The upstream agent must produce a non-empty array result
• Downstream agents reference the original name — fan-out is transparent to them

Result format for gathered fan-out agents:

{
  "response": ["result_0", "result_1", null],
  "status": "partial",
  "fan_out": true,
  "instances": [
    {"index": 0, "status": "success"},
    {"index": 1, "status": "success"},
    {"index": 2, "status": "failed", "error": "timeout"}
  ]
}

Status is "success" (all instances passed), "partial" (some failed), or "failed" (all failed). Partial results are preserved — the ensemble continues with whatever succeeded.

Ensemble Agents (Composable Ensembles)

Agents can reference and execute other ensembles, enabling hierarchical composition:

# child ensemble: topic-analysis.yaml
name: topic-analysis
agents:
  - name: analyst
    model_profile: ollama-gemma-small
    system_prompt: "Analyze the given topic in 2-3 sentences."

# parent ensemble
agents:
  - name: classifier
    script: scripts/classifier.py

  - name: topic-analyst
    ensemble: topic-analysis          # references child ensemble
    depends_on: [classifier]

  - name: synthesizer
    model_profile: default-claude
    depends_on: [topic-analyst]

How it works:

1. The ensemble field identifies which ensemble to execute (resolved by name from .llm-orc/ensembles/)
2. The child ensemble runs as a self-contained execution with its own phases and agents
3. Child executors share immutable infrastructure (config, credentials, model factory) but isolate mutable state
4. Nesting depth is limited (default: 5) to prevent unbounded recursion
5. Cross-ensemble cycles are detected at load time

Input Key Routing

Agents can select a specific key from upstream JSON output using input_key, enabling routing patterns where a classifier produces keyed output and downstream agents each consume their slice:

agents:
  # Classifier produces: {"pdfs": ["a.pdf", "b.pdf"], "audio": ["c.mp3"]}
  - name: classifier
    script: scripts/classifier.py

  # Selects only the "pdfs" array from classifier output
  - name: pdf-processor
    ensemble: pdf-pipeline
    depends_on: [classifier]
    input_key: pdfs
    fan_out: true

  # Selects only the "audio" array
  - name: audio-processor
    ensemble: audio-pipeline
    depends_on: [classifier]
    input_key: audio
    fan_out: true

  - name: synthesizer
    model_profile: default-claude
    depends_on: [pdf-processor, audio-processor]

Behavior:

• input_key selects output[key] from the first entry in depends_on
• If the key is missing or the upstream output is not JSON/dict, the agent receives a runtime error
• Without input_key, the agent receives the full upstream output (backward compatible)
• Composes naturally with fan_out: input_key selects the array, fan_out expands per item
• Works with all agent types: LLM, script, and ensemble

Configuration Status Checking

LLM Orchestra provides visual status checking to quickly see which configurations are ready to use:

# Check all configurations with visual indicators
llm-orc config check

Visual Indicators:

• 🟢 Ready to use - Profile/provider is properly configured and available
• 🟥 Needs setup - Profile references unavailable provider or missing authentication

Provider Availability Detection:

• Authenticated providers - Checks for valid API credentials
• Ollama service - Tests connection to local Ollama instance (localhost:11434)
• Configuration validation - Verifies model profiles reference available providers

Example Output:

Configuration Status Legend:
🟢 Ready to use    🟥 Needs setup

=== Global Configuration Status ===
📁 Model Profiles:
🟢 local-free (llama3 via ollama)
🟢 quality (claude-sonnet-4 via anthropic-claude-pro-max)  
🟥 high-context (claude-3-5-sonnet via anthropic-api)

🌐 Available Providers: anthropic-claude-pro-max, ollama

=== Local Configuration Status: My Project ===
📁 Model Profiles:
🟢 security-auditor (llama3 via ollama)
🟢 senior-reviewer (claude-sonnet-4 via anthropic-claude-pro-max)

Configuration Reset Commands

LLM Orchestra provides safe configuration reset with backup and selective retention options:

# Reset global configuration (safe defaults)
llm-orc config reset-global        # Creates backup, preserves authentication

# Reset local configuration (safe defaults)  
llm-orc config reset-local         # Creates backup, preserves ensembles

# Advanced reset options
llm-orc config reset-global --no-backup --reset-auth           # Complete global reset
llm-orc config reset-local --reset-ensembles --no-backup       # Complete local reset
llm-orc config reset-local --project-name "My Project"         # Set project name

Safety Features:

• Automatic backups - Creates timestamped .backup directories by default
• Authentication preservation - Keeps API keys and credentials safe by default
• Ensemble retention - Preserves local ensembles by default
• Confirmation prompts - Prevents accidental data loss

Available Options:

Global Reset:

• --backup/--no-backup - Create backup before reset (default: backup)
• --preserve-auth/--reset-auth - Keep authentication (default: preserve)

Local Reset:

• --backup/--no-backup - Create backup before reset (default: backup)
• --preserve-ensembles/--reset-ensembles - Keep ensembles (default: preserve)
• --project-name - Set project name (defaults to directory name)

Configuration Hierarchy

LLM Orchestra follows a configuration hierarchy:

1. Local project configuration (.llm-orc/ in current directory)
2. Global user configuration (~/.config/llm-orc/)
3. Command-line options (highest priority)

Library Path Configuration

Control where llm-orc init finds primitive scripts using environment variables or project-specific configuration:

# Option 1: Custom library location via environment variable
export LLM_ORC_LIBRARY_PATH="/path/to/your/custom-library"
llm-orc init

# Option 2: Project-specific configuration via .llm-orc/.env
mkdir -p .llm-orc
echo 'LLM_ORC_LIBRARY_PATH=/path/to/your/custom-library' > .llm-orc/.env
llm-orc init

# Option 3: Use local submodule (development default)
export LLM_ORC_LIBRARY_SOURCE=local
llm-orc init

# Option 4: Auto-detect library in current directory (no configuration needed)
# Looks for: ./llm-orchestra-library/scripts/primitives/
llm-orc init

Priority order:

1. LLM_ORC_LIBRARY_PATH environment variable - Explicit custom location (highest priority)
2. .llm-orc/.env file - Project-specific configuration
3. LLM_ORC_LIBRARY_SOURCE=local - Package submodule
4. ./llm-orchestra-library/ - Current working directory auto-detection
5. No scripts installed (graceful fallback)

Note: Environment variables always take precedence over .env file settings, allowing temporary overrides without modifying project files.

This allows developers to maintain their own script libraries while still using llm-orc's orchestration features.

XDG Base Directory Support

Configurations follow the XDG Base Directory specification:

• Global config: ~/.config/llm-orc/ (or $XDG_CONFIG_HOME/llm-orc/)
• Automatic migration from old ~/.llm-orc/ location

Cost Optimization

• Local models (free) for systematic analysis tasks
• Cloud models (paid) reserved for strategic insights
• Usage tracking shows exactly what each analysis costs
• Intelligent routing based on task complexity

Development

# Run tests
uv run pytest

# Run linting and formatting
uv run ruff check .
uv run ruff format --check .

# Type checking
uv run mypy src/llm_orc

Research

This project includes comparative analysis of multi-agent vs single-agent approaches. See docs/ensemble_vs_single_agent_analysis.md for detailed findings.

The short version: orchestrated multi-agent systems maintain accuracy at scale where single-agent approaches collapse. Mixture-of-Agents ensembles of open-source models have matched or exceeded frontier model performance on established benchmarks. Cascade routing strategies have replicated frontier quality at a fraction of the cost. The evidence supports the architectural bet this project is built on.

Philosophy

Coordination over scale. Process over generation.

The concentrated AI buildout optimizes for one thing: making the generative phase faster. But generation was never the bottleneck. Evaluation — the human judgment that determines whether output is correct, appropriate, and worth shipping — is the binding constraint. Faster generation without proportionally better evaluation just produces more to review.

LLM Orchestra takes the opposite position. An ensemble of smaller, specialized models — each focused on a bounded analytical task — produces structured output designed for human evaluation. The system decomposes problems so that each agent owns its contribution. A security reviewer finds vulnerabilities. A performance analyst identifies bottlenecks. A synthesis agent integrates their findings. The human evaluates a structured analysis, not raw generation.

Running models locally is a practical choice, not an ideological one. No per-query billing means you can run systematic analysis across an entire codebase without watching a meter. Data stays on your hardware. And the local/cloud mix lets you put cost where it matters — expensive models for strategic insight, free local models for systematic coverage.

License

AGPL-3.0 License - see LICENSE for details.