What problem does this solve?

If you've used AI for coding, you know the drill: every new session the assistant has no idea about your project. You explain the same things over and over.

Rudder fixes that. Your specs, memory, and task state all live in the repo. Every session gets exactly the context it needs — automatically.

| Capability | What it changes | | --- | --- | | Auto-injected specs | Write your conventions once in .rudder/spec/, Rudder injects the right files into each session. No more copy-pasting rules every time. | | Task-centered flow | PRDs, implementation records, and review context all live in .rudder/tasks/. AI work stays structured, not scattered across chat history. | | Project memory | .rudder/workspace/ journals carry forward what happened last time. New session? It knows where you left off. | | Team-shared standards | Specs version-control with your code. One person figures out a good pattern, everyone gets it for free. | | 14 platforms, one setup | Same Rudder structure everywhere. Switch tools, keep your workflow. |

Supported Platforms

14 AI coding platforms, one setup. Pick what you use:

| Agent | IDE / Platform | | --- | --- | | Claude Code | CLI, VS Code, JetBrains, Web | | Gemini | Gemini CLI | | Cursor | Cursor IDE | | Codex | VS Code | | OpenCode | OpenCode IDE | | CodeBuddy | CodeBuddy IDE | | Qoder | Qoder IDE | | Kiro | Kiro IDE | | Pi | Pi CLI | | Windsurf | Windsurf IDE | | Copilot | VS Code, JetBrains | | Antigravity | Antigravity CLI | | Kilo | Kilo CLI | | Droid | Factory IDE |

What you need

• Node.js >= 18
• Python >= 3.9

Quick Start

# 1. Install
npm install -g @mengde1231/rudder@latest

# 2. Initialize (all platforms)
rudder init -u your-name

# 3. Or just the ones you actually use (cleaner)
rudder init --cursor --opencode --codex -u your-name

How it works

Rudder runs a 4-phase loop every session:

1. Plan — rudder-brainstorm walks through requirements and writes prd.md. Research-heavy stuff goes to a rudder-research sub-agent. Output: curated specs + research files, wired up via implement.jsonl / check.jsonl.
2. Implement — rudder-implement writes code from the PRD with context already injected. No git commit here — that's intentional.
3. Verify — rudder-check reviews the diff against specs, runs lint, type-check, compile, and tests. Fixes what it can, reports what it can't.
4. Finish — rudder-update-spec promotes new learnings back into .rudder/spec/ so the next session starts smarter.

FAQ

Those are useful entry points — until they become 3,000-line monoliths nobody reads. Rudder layers things: scoped specs, per-task PRDs, workflow gates, cross-platform adapters. One simple rule: don't cram everything into one file.

Nope. Rudder is a project-layer thing. Works across 14 coding agents and IDEs. Use Gemini for frontend today, Claude Code for backend tomorrow, Codex for review the day after.

Both. Solo devs get project memory and repeatable workflow. Teams get the bigger win: shared standards, clear task boundaries, reviewable context, and platform portability.

Nope. Most teams let AI draft specs from existing code first, then tighten the important stuff by hand. Rudder works best when the high-signal rules are explicit and versioned — the rest AI can figure out.

Nah. Personal workspace journals are per-developer. Shared specs and tasks go through git like any other artifact — merge conflicts are just merge conflicts, nothing new.

Acknowledgments

Rudder wouldn't exist without the open-source coding agent community. We studied how many projects solve the same problems — context injection, workflow gates, multi-platform support — and built the version we wished we had.

Star History