$ kora

Many agents.
One cohesive result.

Specialized AI agents that research, plan, build, review, and validate your code until it's right. Named after the kora, a 21-string instrument where one musician plays every string.

$ npm install -g @usekora/kora
Works with
Philosophy

The safety net that AI agents never had

Human developers have code review, security audits, and QA. Why should AI agents skip all of that?

01

Describe the task

Interactive session or one-shot command. Plain English. Kora auto-classifies complexity and scales the pipeline to match.

02

Agents challenge each other

The plan is reviewed, security-audited, and judged before a single line of code is written. Real issues survive. Nitpicks don't.

03

Ship with confidence

Parallel implementors write code in isolated worktrees. Every diff is reviewed. A Validator confirms the result matches the intent.

The Pipeline

Every stage has a purpose.
Every agent has a role.

Two quality loops. A judge at each gate. The plan is stress-tested before code exists, and every diff is audited after.

Stage 1 / Research
Explore & propose
The Researcher maps your codebase, clarifies intent with you, and drafts a detailed plan.
Researcher
Stage 2 / Review Loop
Challenge the plan
Reviewers find gaps. A Security Auditor checks for vulnerabilities. The Judge filters noise so only real issues go back for revision.
Plan Reviewer Security Auditor Judge
Stage 3 / Plan & Build
Parallel implementation
Planner breaks work into parallel tasks. Test Architect designs test strategy. Implementors execute simultaneously in isolated worktrees.
Planner Test Architect Implementors
Stage 4 / Audit & Validate
Every diff, reviewed
Code Reviewer and Security Auditor inspect every change. The Validator confirms the implementation matches the original intent.
Code Reviewer Code Security Validator
Researcher
Reviewer Security
Judge
Planner Tests Fleet
Review Audit
Validate
Features

Built for how you actually work

Adaptive

Pipeline scales to the task

A typo fix skips straight to implementation. A complex feature gets the full pipeline with security auditing. The Researcher auto-classifies so you never have to.

Routing

Smart provider routing

Presets route each agent to the best available CLI tool. Claude for deep reasoning, Codex for fast validation, Gemini for large-context research.

Agnostic

No API keys, no lock-in

Kora spawns your existing CLI tools as subprocesses. Install one, two, or all three providers. Authentication is theirs, not ours.

Parallel

Simultaneous execution

A 4-task feature gets 4 agents writing code at once, each in its own isolated git worktree. Merge when done.

Resumable

Ctrl+C and come back

Every stage transition is saved to disk. kora resume picks up exactly where you left off. Nothing is lost.

Control

You stay in the loop

Configurable checkpoints let you approve at any stage. Remote operations always require explicit consent, even in --yolo mode.

In Action

What a run looks like

kora
researcher .............................. analyzing Found 47 files relevant to your request. Proposing approach with 3 key changes... --------------------------------------------------------------- reviewer ................................ analyzing plan HIGH No database migration strategy MED Missing error boundary LOW Could use const enum -- dismissed judge ................................... evaluating DB migration accepted Error boundary accepted Const enum dismissed plan approved --------------------------------------------------------------- implementing ............................ 2 of 4 T1 claude ############ done 34s 7 files T2 codex ############ done 12s 3 files T3 claude ##########.. running T4 gemini ###......... running --------------------------------------------------------------- implementation complete -- 4 tasks, 1m 23s merged 4 branches
Get Started

Stop generating code.
Start orchestrating it.

Install in seconds. Works immediately with whatever AI CLI tools you already have.