grex

Grex architecture

Overview

Hand a bounded task to AI agents on computers you own. A second agent reviews the exact result against your acceptance criteria before anything is accepted. Grex has two parts: a small control plane and the machines you enroll.

Control plane

The control plane stores missions, sends assignments to your machines and keeps the evidence. It runs no AI and holds no Git credentials. Accepted code is pushed as a pull request from the operator's own machine.

Nidi

Nidi are the machines you enroll to do the work. Today these are Linux and Windows computers. Each one can run several missions at once, and Grex shows its disk, memory and CPU health.

Runtimes and models

Grex uses the agent command-line tools you install and sign in to yourself: Codex with a ChatGPT plan, and Claude Code with a Claude plan. It can also call the MiniMax, Anthropic and Gemini APIs, which are billed per use, or a local model through Ollama. The builder and reviewer are chosen separately, so a cheaper model can build while a stronger one reviews.

Mission lifecycle

A mission has an objective, acceptance criteria and limits: the files it may change, a starting commit and a deadline. A builder agent does the work in a sandbox. Mechanical checks, such as the build and tests, run next, and the patch, logs and check results are kept. A separate reviewer agent judges the exact submission against each criterion. You decide: accepted code becomes a pull request, and a rejected mission can be corrected with the reviewer's reasons attached.

Research briefs and site reports

Research briefs are cited, and their quotes and dates are checked mechanically. Site reports use a browser to check a live site and report faults. Both can run on a daily schedule, and a schedule pauses itself after repeated failures.

Operating Grex

The dashboard, the HTTP API and MCP do the same things. Work is grouped into lines: operate, grow, evolve and discover.

Terms

  • Grex: the whole system, and the brand.
  • Nidus (plural Nidi): a machine you enroll to do the work.
  • Manipulus (plural Manipuli): a standing team for one line of work, made of its agents, mission templates and schedules.
  • Duoviri: the two-agent rule. A builder does the work, and a separate reviewer decides whether it meets the criteria.