AI Orchestration Framework
IRA Framework is an orchestration layer that runs Claude Code, Codex, and Gemini through a 5-stage software pipeline — brainstorm, specs, implement, audit, document — pausing at human approval gates so nothing ships without your sign-off. It works directly in your own repository, with no API key or database required.
Runs in your repo · Uses your Claude Code login · No API key · No database
The short version
IRA Framework is an orchestration layer for AI coding agents. It runs your work through a 5-stage software lifecycle — brainstorm, break into specs, implement, audit, then document — and hands you a human approval gate at each high-leverage step, so nothing ships without your sign-off. The AI does the work inside your own repository; you stay in command of what ships.
IRA ships 4 pipelines — Full, Bug, Docs, and Improve — and a feature build passes through 2 approval gates before anything merges. Every stage is routed to the right model across 3 AI coding CLIs: Claude Code, Codex, and Gemini. After each run, IRA compares what it built against the spec and your decisions, then proposes concrete improvements to how it works — so it gets better at your projects the more you run it.
Core capabilities
Six things that turn an AI coding tool into a system you can trust with real work.
Full feature builds, bug fixes, docs, and improvements each run as a structured pipeline — not a single prompt. Runs proceed on their own but stop at approval gates where it matters, so nothing ships without your yes. Reject with a reason and IRA adjusts.
› Approve / Reject at each gate · optional typed "I approve"
Reasoning stages run on deep models; coding stages run on fast ones — automatically. Route any stage to Claude Code, Codex, or Gemini, and mix providers within a single pipeline. Reasoning runs read-only; coding runs with full access inside the target repo.
› Per-stage provider + model routing · Opus reasons, Sonnet codes
IRA records what it learns about each project into a knowledge graph, and after every run a review pass compares the result against the spec and your decisions, then drafts concrete config improvements. You approve them. The framework gets better at your projects the more you run it.
› Post-run analyzer → proposals you approve · versioned & reversible
Encode how you work once — custom sub-agents, markdown rules, and reusable skills — scoped to the exact pipelines and stages where they should fire, then injected automatically. Edit pipelines visually on a draggable node canvas. Every change is versioned and snapshotted.
› Scoped sub-agents · rules · skills · editable node-graph pipelines
IRA interviews you to build a profile of your role, your tools, and how you like to communicate — then folds it into every run so the work and the tone fit you. Have more than one job? It can propose a tailored pipeline per role. You don't have to be an engineer to be understood.
› Built by an AI interview · steers every pipeline and reply
A live AI assistant sits in a side dock on every page. Describe the change you want — "add an agent that runs the linter after implement" — and it applies real config edits as you watch, one card per change. Changed your mind? Undo the whole turn in a click.
› Natural-language config · live edit cards · instant Undo
Under the hood
Audits surface a prioritized backlog; run non-conflicting items at once, each in its own git worktree, auto-merged back.
A React dashboard streams every run live; chats and jobs keep running when you navigate away and notify you when done.
Project knowledge is captured as a graphify graph and can export to an Obsidian vault you can browse.
No database, no queue. Every run lives under your project in a .ira/ folder — inspectable, portable, git-friendly.
Ask questions about any project and get answers grounded in its graph — with inline, zoomable Mermaid diagrams.
Export your whole setup — agents, rules, pipelines, skills, profile — to one file and replicate it on another machine.
How it works
Pick a pipeline, give it a goal and a target repo, and watch it work. Approve the gates. Everything runs against your codebase; nothing hides in a black box.
| Pipeline | What it's for | Stages |
|---|---|---|
| Full | Build a new feature |
Brainstorm→
Breakdown Specs→
Gate→
Implement→
Audit→
Gate→
Document
|
| Bug | Fix a bug |
Reproduce→
Root Cause→
Gate→
Fix→
Verify→
Gate→
Document
|
| Docs | Write / update docs |
Survey→
Draft→
Gate→
Finalize
|
| Improve | Improve an existing area |
Brainstorm→
Breakdown Specs→
Gate→
Implement→
Audit→
Gate→
Document
|
Every run is written under your target project: <your-repo>/.ira/runs/<id>/ — run state, per-stage artifacts, and full logs. Filesystem is the source of truth.
Who it's for
IRA gives everyone the same rigor — and meets each person where they are.
Offload the grind — reproduce, root-cause, implement, audit — without giving up the wheel. Every diff runs in your own repo, gated by you, with the reasoning and the tests in plain view. Encode your conventions as rules and agents once and reuse them everywhere.
You don't have to be a 10x engineer. Describe the outcome and approve the milestones — the "Me" profile and plain-English config mean IRA speaks your language, not just the terminal's. Ship real features while staying in control of what goes live.
Standardize how work gets built across every project — one set of agents, rules, and skills, consistently applied. Turn an audit's backlog into parallel group runs and ship a whole batch unattended, each task isolated in its own branch and merged cleanly.
The learning loop
When a run finishes, IRA reviews its own work against the spec and your gate decisions, then proposes specific improvements to its agents, rules, and pipelines. You approve what you like. Nothing changes without you — and every change is versioned and reversible.
run→review→proposal→your approval→sharper next time
Access
Sign in to your IRA workspace to connect your repositories and run your first pipeline.
FAQ
IRA Framework is an orchestration layer that drives AI coding agents — Claude Code, Codex, or Gemini — through structured software pipelines with human approval gates. It plans, implements, audits, and documents work inside your own repository, records what it learns into a knowledge graph, and proposes improvements to its own configuration after every run.
No. Developers get fine-grained control, but IRA is built for founders and solo builders too. You describe the outcome and approve the milestones; a "Me" profile and a plain-English configuration assistant let you steer the framework without living in the terminal.
Claude Code is the primary, default, and most capable provider. Codex and Gemini are also supported, and you can route each pipeline stage to a different provider and model — deep-reasoning models for planning and review, fast models for coding — even within a single pipeline.
IRA runs on your machine and works inside your own repository. It uses your existing Claude Code login, so no separate API key is needed. Run state, artifacts, and logs are written to a .ira/ folder under your project — the filesystem is the source of truth, with no external database.
Gates are the points where a running pipeline pauses for your decision. When a run reaches a gate it shows "awaiting approval"; you Approve to continue or Reject with a reason, and IRA adjusts. You can require typing "I approve" for an extra confirmation step.
Yes. IRA includes a Workflows layer — custom sub-agents, markdown rules, and reusable skills, each scoped to the pipelines and stages where it should apply. Pipelines are fully editable on a visual node canvas, and every change is versioned and reversible.
Yes. Point it at any local repository. IRA builds a knowledge graph of the project, can extract an architecture document from it, and grounds its planning and its project chat in that understanding of your actual code.
After each run, a review pass compares the result against the spec and your gate decisions, then drafts concrete improvements to IRA's agents, rules, and pipelines. You approve the ones you want. Because the managed config is versioned and snapshotted, every change is auditable and reversible.
Structured pipelines, the right model for every step, and a framework that learns your project. That's IRA.
Sign in