Let Claude Code and Codex collaborate on your large-scale coding tasks. Decompose goals into DAGs, execute in parallel, converge on verified results.
Large tasks break single AI sessions. You end up manually stitching results together.
Refactoring 15 files? A single agent loses track of changes made earlier. Inconsistent edits, missed dependencies, broken imports.
Independent tasks run sequentially. 100 file fixes that could take 2 minutes instead take 30. No retry on transient failures.
Planning needs reasoning. Execution needs speed. Using one model for everything means compromising on both.
You become the orchestrator — running agents, checking outputs, feeding results to the next step. That's not automation.
One command. Automatic decomposition, parallel execution, cross-agent verification.
# One command replaces hours of manual coordination
$ mo do "Add JWT auth with middleware, routes, tests, and docs"
# Decomposing goal into 6 tasks...
# Task 1: auth-middleware (claude) ─── running
# Task 2: token-utils (codex) ─── running
# Task 3: auth-routes (codex) ─── waiting [depends: 1,2]
# Task 4: auth-tests (codex) ─── waiting [depends: 1,3]
# Task 5: api-docs (codex) ─── waiting [depends: 3]
# Task 6: integration-review (claude) ─── waiting [depends: 4,5]
# All 6 tasks completed. 0 failures. Duration: 4m 23s
| Capability | Single Agent | Master Orchestrator |
|---|---|---|
| Multi-file refactoring | Context overflow, inconsistent | DAG decomposition, parallel |
| Bulk operations (100+ files) | Sequential, no retry | 16 parallel workers, auto-retry |
| Error recovery | Manual restart | Classification, backoff, fallback |
| Provider flexibility | Stuck with one model | Mix Claude + Codex per phase |
| Convergence detection | None | Plateau, regression, deterioration |
| Crash recovery | Start over | Checkpoint + resume |
Goals are automatically broken into dependent sub-tasks. Independent tasks run in parallel.
Route each phase to the best AI agent. Claude for reasoning, Codex for code generation.
High-throughput bulk execution with 16 parallel workers, auto-retry, and syntax validation.
Classifies errors (rate limit, context overflow, transient) and applies the right recovery strategy.
Monitors for plateaus, regressions, and quality deterioration. Automatically escalates or rolls back.
SQLite-backed checkpoints. Resume any interrupted run from where it left off.
Per-provider budget limits with accounting mode. Never exceed your spend ceiling.
The orchestrator can analyze its own runs and propose improvements to your workflow.
TOML-based configuration for providers, routing, retry policies, and validation rules.
# Install
$ git clone https://github.com/amber132/Master-Orchestrator.git
$ cd Master-Orchestrator
$ pip install -e ".[dev]"
# Run your first orchestrated task
$ mo do "Add input validation to all POST endpoints"