Open source · Self-hosted

Open-source AI platforms for your team, running on your own infrastructure

GreenOnGrey builds two products. Nabu gives every employee a personal AI agent. Hammurapi runs the product development cycle on Nabu, with its agents as partners.

GreenOnGrey logo: a blade of grass growing through a cracked grey stone

Principles

Open source

Apache 2.0. The code and the documentation are public, including the limits we know about.

On your infrastructure

Both products run on your servers, with the models you choose. Nothing is hosted by us.

People decide

Agents do the work between decisions. Approvals, signatures and releases stay with people, and every tool call is audited.

early stage

Nabu

An on-premises platform of AI agents: a personal agent for every employee and service agents for your products.

Nabu is the Mesopotamian god of writing and scribes.

A personal agent for every employee

Its own memory, files and scheduled tasks. Each user's agent and sandbox run in their own pods, and connections use the user's own credentials.

Guardrails and a sandbox

Commands run in a sandbox with no capabilities and no cluster token. Administrators choose the models, the tools and the catalog; read-only connections expose only reading tools; every tool call is in the audit trail.

Built into your products

A product starts a service agent through an API or acts on behalf of a user. For code, the agent works in a sandbox on a copy the product uploads and returns the changed files; the sandbox has no git credentials.

One agent across channels

The web, mail, Telegram, VK Teams and the chat of Hammurapi lead to the same main conversation and the same memory. A group chat gets an agent of its own.

early stage

Hammurapi

A web platform that runs the product development cycle with the agents of Nabu as partners, from an idea to the value metric after release. It works only with Nabu; git stays the source of truth.

Named after Hammurabi, the Babylonian king known for his code of laws.

  1. Discovery

    • An idea or a problem becomes an issue
    • The agent works out its value and a metric
    • An expert accepts it into a feature
  2. Development

    • Specifications pass quality gates in order
    • The agent writes the tech and QA specifications
    • Code arrives as pull requests in every service
  3. Delivery

    • Merge and deploy
    • Feature flag and the value metric
    • Confirmation or one-click rollback

What a release teaches becomes the input of the next Discovery

More than spec-driven: every change is traced

Requirements have IDs, and a coverage matrix links them to services, test cases and pull requests. Every commit names its feature, task and initiator. Any edit returns a specification to draft, so an approval always covers the current text.

From research to production

One cycle covers the analysis of an issue, five specifications, code in every service, validation against CI results, the release and the metric that confirms its value.

People sign, the agent works

People accept issues, approve specifications, sign the validation and confirm releases. The agent acts within the user's permissions, has no delete tool and writes code in a sandbox of Nabu without git credentials.

Hammurapi runs on Nabu

Hammurapi has no agent of its own and does not start without Nabu. It uses Nabu in two ways.

Hammurapiproduct development cycle
  • The chat in Hammurapi is a window to the employee's personal agent in Nabu: the same conversation, memory and files.
  • The scenarios of the cycle, from issue analysis to code generation, run on Nabu service agents; code is written in a sandbox of Nabu.
Nabuplatform of AI agents
  • LLMs
  • SKILLS
  • MCPs

Nabu gives agents the models, skills and MCP servers that your administrators connect.

In turn, the personal agent reads and creates issues in Hammurapi on behalf of its user.

Install on your server

Nabu, or Nabu with Hammurapi, with one command: on Debian, Ubuntu or RHEL with minikube, or into your Kubernetes cluster.

  1. Run the command

    It downloads the latest release of the installer and opens its wizard. Each release is signed and has a checksum: the guide shows how to verify them.

  2. Answer the wizard

    The domain, how people sign in, built-in or your own Postgres, Kafka and S3, and MCP servers. Nothing changes until you confirm.

  3. Open the products

    The installer checks the install and prints the addresses. An update is one command too.

curl -fsSL https://github.com/GreenOnGrey/installer/releases/latest/download/installer.tar.gz | tar xz && ./installer/installer

What's next

A direction, not a feature list. None of this is in the products yet.

planned

Nabu

One place to analyse an employee's work communications: mail, messengers and the company's products in one agent.

planned

Hammurapi

The platform watches how the product works technically and proposes changes as new issues in Discovery.

Both products are at an early stage

Read the code, install the products, open an issue and tell us what breaks. A star helps others find the projects.