ralph-o-matic
A job queue server that runs iterative AI coding refinement loops — locally via Ollama, against the Anthropic API, or through OpenRouter — so you can queue work, walk away, and review results as PRs.
The Problem
Iterating on code until tests pass and acceptance criteria are met produces excellent results, but doing it manually is tedious and doing it against cloud APIs burns credits fast. Most iterations are mechanical refinement, not creative problem-solving.
The Solution
Draft your implementation with Claude Code + Opus. Then hand off to ralph-o-matic, which runs the refinement loop with built-in circuit breakers, retry logic, session continuity, and per-iteration commits. Use local models via Ollama to save API credits, use the Anthropic API when you need Claude-grade quality, or use OpenRouter for access to models from multiple providers on a pay-per-token basis.
Your Dev Env (Opus 4.6) Ralph-o-Matic Server
┌──────────────────────┐ ┌─────────────────────────┐
│ /feature-pipeline │ │ ralph-o-matic-server │
│ │ │ │
│ 1. Brainstorm spec │ ◄─ you │ │
│ 2. Brainstorm design │ ◄─ you │ │
│ ── walk away ── │ │ │
│ 3. Align designs │ auto │ │
│ 4. Write plans │ auto │ │
│ 5. Align plans │ auto │ │
│ 6. Implement │ auto │ │
│ 7. Generate loop │ auto │ │
│ 8. Submit to ralph ──│───────────│─► Queue → Execute loop │
│ │ │ Commit → Push → PR │
└──────────────────────┘ │ 9. Post-completion hook │
│ → Auto PR review │
│ → Teams notification │
└─────────────────────────┘
Features
- Three backends — run loops against local Ollama models, the Anthropic API (via Claude Code), or OpenRouter (Kimi, Grok, Devstral, and more)
- Job queue with priority scheduling (high/normal/low), pause/resume, drag-and-drop reordering
- Circuit breaker — detects no-progress loops and repeated errors, stops wasting compute
- Session continuity — resumes Claude sessions across iterations for better context
- Per-iteration commits — every successful iteration is committed, so crashes don't lose work
- Retry with backoff — transient errors are retried with exponential backoff
- Rate limiting — configurable per-hour call limits (essential for Anthropic API, disabled for Ollama)
- Smart model selection — detects your hardware (RAM, GPU VRAM, Apple Silicon) and recommends optimal model placement
- Split-device inference — run the large model on CPU/RAM and the small model on GPU, or both on GPU if you have the VRAM
- Remote Ollama support — point at a remote Ollama instance instead of running locally
- Feature pipeline —
/feature-pipeline skill automates the entire workflow: brainstorm spec, brainstorm design, align, plan, implement, submit to ralph, auto PR review. You interact during brainstorming, then walk away.
- Post-completion hooks — run a shell command when jobs finish (e.g., trigger automated PR review via Claude Code)
- Notifications — email (SMTP) and Microsoft Teams webhook notifications on job completion, failure, or cancellation. Skills can send arbitrary messages via
ralph-o-matic notify.
- Authentication — optional Microsoft Entra ID (Azure AD) SSO with role-based access control
- Web dashboard with live updates via SSE
- Git integration — auto-clones repos, creates result branches, opens PRs on completion
- Claude Code skills —
/feature-pipeline for fully automated end-to-end workflows, plus composable sub-skills (spec-to-design, auto-ralph-prep, auto-ralph-submit) and interactive skills (brainstorm-to-ralph, direct-to-ralph, plan-to-ralph)
- Cross-platform — macOS, Linux, and Windows; amd64 and arm64
Quick Start
Install (macOS / Linux)
curl -fsSL https://raw.githubusercontent.com/dbinky/ralph-o-matic/main/scripts/install.sh | bash
Install (Windows PowerShell)
irm https://raw.githubusercontent.com/dbinky/ralph-o-matic/main/scripts/install.ps1 | iex
Both installers detect your hardware, let you choose a backend, recommend models, install dependencies, and start the server. Use --yes (bash) or -Yes (PowerShell) for non-interactive mode.
Installation Modes
dbinky