Case study

Building Zixt, an AI-teammate platform

Zixt gives a business AI teammates with real roles, coordinated by an AI Manager, working on the company's own machines and accounts. GreenCandle built the whole product and runs it today.

The problem

An AI that works for a business needs more than a chat window.

It needs a role and a brief. Access to the company's tools without every password.

A way to ask a person when it should. A record of what it did.

A Manager that turns requests into tasks and reports back with receipts. Most of that is platform engineering, not prompting.

The teammates Zixt gives a business
  • AtlasWebsite
  • MayaMarketing
  • SamSupport
  • JordanSales
  • RileyOffice
  • TheoDeveloper
Each teammate owns one area, works every day, and knows what needs a person's sign-off.

What we built

Six systems, one product.

Cloud control plane

TypeScript on Bun: organizations, agents, orchestration, webhooks, billing, a communications gateway for Slack and email, and a WebSocket gateway to every machine.

Desktop runtime

A headless Node runtime that runs coding agents on the customer’s own subscription and executes every tool call locally. Inference keys and data never leave the machine.

Credential custody

Secrets live encrypted in the cloud vault. A machine receives a grant scoped to one task, in memory, for as long as that work runs.

Guardrails

Instructions first: the policy is rendered into each session, autonomy is the default, and the teammate asks a person when the policy says so.

Integrations and voice

GitHub, Slack, Linear, email, WhatsApp, a browser teammates can drive when there is no API, voice mode, and phone calls.

Web app and installer

A React app for people, usable at phone width, and a signed, notarized companion installer for Windows and macOS that keeps the runtime updated.

How it is built

AI-native development, under a gate that does not bend.

Agents write and review much of the code. Senior engineers own the architecture, the security model, and every release. The gate is what makes that safe.

  • Integration tests against a real database and real WebSockets
  • Property-based tests for the task engine
  • Browser journeys against the production bundle
  • An eight-lane CI gate on a self-hosted fleet, under seven minutes, no retries
  • Automatic releases to the package registry, the installers, and the cloud
Connected toLive
GitHubSlackLinearGmailWhatsAppMCPAnthropicOpenAIMongoDB
In private development at zixt.ai. Every service on this site is a piece of the engineering that made it.

Your project

Building something with the same shape?

Agents, control planes, desktop runtimes, integrations, or the pipeline behind them.

Tell us about your project
  • The same teamThe engineers who built Zixt work on client projects.
  • The same gateReal tests, browser journeys, and CI from the first commit.