OpenJunto
Coding agents agree too easily.
OpenJunto makes them argue.
Today: Claude Code plugin only.
A coordinated team of expert sub-agents critiques each result instead of agreeing with it. Slower, costs ~2–5× the tokens — the cost of catching what you'd otherwise ship.
The review loop
- Request "Add rate limiting."
- Experts Security · DevOps · Engineer
- Adversarial review a different expert challenges the work STRONGEST OBJECTION FALSIFIER
- Synthesis findings + dissent, handed back
↺ unresolved objections loop back to the experts
What it is
OpenJunto adds peer review to AI coding agents. Today it ships as a Claude Code plugin. A Manager agent triages each request, pulls in the relevant domain experts, and routes the work through mandatory peer review — so a different expert challenges each result instead of agreeing with it.
The point is the review. A single agent's output looks coherent because nothing challenged it. OpenJunto adds a second look: it tests failure modes, names the strongest counterargument, and catches what the first pass missed.
Ask it to add rate limiting and it won't just write middleware — Security flags the bypass vectors, DevOps sizes the failure modes, and a reviewer probes both.
Built for senior engineers and tech leads who want a structured second opinion on critical work. It's slower and uses ~2–5× the tokens of a single agent — the cost of catching what you'd otherwise ship. See WHY.md for the honest version, including when it's not worth it.
- Strongest objection & falsifier
- Every recommendation carries its best counterargument and the condition under which it would break in production.
- Complexity triage
- Simple / Moderate / Complex tiers scale the process to the task.
- Adversarial review
- All Moderate and Complex work is reviewed by an expert from a different domain.
- Circuit breakers
- Work auto-escalates to you after 3 revision cycles or 2 hours without meaningful progress.
Supported interfaces
OpenJunto's coordination layer is defined once and generated per interface. Today, Claude Code is the only shipping target.
| Interface | Status | Install |
|---|---|---|
| Claude Code | Shipping (pre-1.0) | /plugin install oj@openjunto |
No other interfaces ship today. An OpenAI Codex target (oj-codex) is now generated and in beta — its bindings are validated against the Codex docs, but portability has not yet been demonstrated end-to-end on a live Codex CLI, so it isn't a shipping target yet.
Get started on Claude Code
The plugin lives in oj-claude. Install it from the marketplace inside a Claude Code session:
/plugin marketplace add openjunto/oj-claude
/plugin install oj@openjunto
Then start a new session to load it. Put it to work with /oj:cycle <what you want done> —
the manager triages, selects the experts, and runs the peer review, surfacing dissent. You don't
name experts; the manager picks them.
Repositories
- oj-claude The Claude Code plugin — generated artifacts and Claude-specific configuration.
- oj-codex The OpenAI Codex plugin — generated, in beta. Bindings validated against the Codex docs; not yet verified end-to-end on a live Codex CLI.
- juntospec The specification corpus — the genome. Defines what to generate; contains no generated code.
- juntogen The generator pipeline that produces oj-claude from juntospec.