🌳

kodama

A portable, provider-neutral agent pack for the Kiro CLI.
One orchestrator. Eight specialists. Zero lock-in.

Get startedView on GitHub

Why?

A single AI agent juggling planning, coding, reviewing, and deploying loses context, makes sloppy mistakes, and has way too many permissions. You end up babysitting it.

Kodama fixes this by splitting the work across specialists — the scout can only read, the smith can only write with approval, the critic only reviews. Each agent does one job well, with the minimum permissions needed.

You describe what you want. Kodama figures out who does what.

Design principles

📦

Portable

No provider lock-in. Works with any model supported by Kiro CLI.

🛡️

Safe by default

Implementation tools require approval. Read-only agents stay read-only.

🧠

Learns your project

Persistent memory across sessions. No repeated discovery.

Evidence-based

Work is reviewed and validated before completion.

The team

Kodama delegates to purpose-built specialists. Each has clear boundaries and minimal permissions.

🌳kodama

Orchestrator

Plans, delegates, reconciles, verifies

🔍kodama-scout

Reconnaissance

Read-only codebase exploration

📚kodama-scholar

Research

External docs, APIs, libraries

🧠kodama-sage

Reasoning

Architecture, debugging, tradeoffs

🎨kodama-artist

UI/UX

Components, styling, accessibility

⚒️kodama-smith

Implementation

Code, tests, refactors

🔬kodama-critic

Review

Independent diff review

🔥kodama-forge

Infrastructure

CI/CD, Docker, IaC, deploy

✍️kodama-scribe

Documentation

READMEs, ADRs, changelogs

Install

One command. Requires kiro-cli and python3.

One-liner install

curl -fsSL https://dl.getkodama.dev | bash

Or directly from GitHub

curl -fsSL https://raw.githubusercontent.com/amergrgic/kodama/main/install-remote.sh | bash

Or clone and run locally

git clone https://github.com/amergrgic/kodama.git
cd kodama && ./setup.sh

Then start a session:

kodama

The alias is added by default. Use --no-alias to skip it.

More than a chat alias

The kodama command is a unified entrypoint.

$ kodama # start a chat session
$ kodama stats # usage insights
$ kodama memory # view project memory
$ kodama doctor # check installation health
$ kodama update # self-update

🧠 Project memory

Kodama remembers what it learns. Build tools, architecture decisions, failed approaches, and team conventions persist across sessions automatically.

Session 1

Scout discovers: vitest, pnpm, React 18, tRPC. Facts are saved automatically.

Session 2+

Kodama already knows your stack. No re-discovery, no wasted context. Faster and more accurate.

In action

Same prompt. Different approach.

Default Kiro agent

You: Create an Express API with /health, tests, Dockerfile, and README.
Writing Express app…
Adding /health route…
Writing tests…
— forgot the Dockerfile —
Writing README…
— README is generic, doesn't mention /health —
— no review, no validation —
⚠ Dockerfile missing
⚠ README incomplete
⚠ No review step
You: "You forgot the Dockerfile…"

Kodama

You: Create an Express API with /health, tests, Dockerfile, and README.
scout Checking workspace…
smith Express app + /health + tests…
forge Multi-stage Dockerfile…
scribe README with API docs…
critic Reviewing all changes…
✓ 4 tests passing
✓ Docker build succeeds
✓ README covers /health endpoint
✓ Critic: no issues
Done. Nothing forgotten.