The open-source platform for digital employees

Build, train, deploy, and operate AI employees. ZevZev turns an AI worker into a deployable, governable, testable software unit — Employee as Code for the AI workforce era.

Get Started Star on GitHub

Kubernetes manages Containers. ZevZev manages Digital Employees.

zevops.employee.yaml

apiVersion: zevzev.io/v1
kind: Employee
metadata:
  name: zevops
  description: AI SRE employee for incident response
spec:
  role:
    title: SRE Engineer
    mission: Monitor systems, triage alerts, propose fixes.
  runtime:
    type: hermes
    model: { provider: openai, name: gpt-5 }
  skills: [incident-response, kubernetes-debug, github-issue-management]
  tools: [github, grafana, kubernetes]
  permission:
    level: proposer
    requireApproval: [production-change, external-message]

// Employee as Code

An AI employee is a deployable, governable, testable unit.

Just like a Kubernetes Deployment describes a workload, a ZevZev Employee manifest declares who the worker is, how it thinks, what it can do, and what it is allowed to do — versioned in Git, reviewed in pull requests, and rolled out with confidence.

spec.role

Identity & mission

The job title and the mission the employee is accountable for — the north star that guides every decision it makes.

spec.runtime

How it thinks

A runtime-neutral binding to Hermes, the OpenAI Agent SDK, LangGraph, or CrewAI, plus the underlying model. Swap it without rewriting the employee.

spec.skills

What it knows

Reusable capability packages composed onto the employee. Skills over prompts — trained, shareable, and independently versioned.

spec.tools

What it can touch

Governed access to external systems through the Tool Gateway, so credentials are brokered — never baked into a prompt.

spec.permission

What it may do

Governance by design: authority levels and human-approval gates for sensitive actions like production changes or external messages.

kind: Employee

One portable file

Everything above lives in a single declarative manifest you can diff, test, and deploy — the contract between humans and a digital worker.

// The platform

Everything you need to run a digital workforce.

A local-first control plane and a composable set of modules — from the CLI in your terminal to the evaluation harness that decides whether an employee can be trusted with a role.

Zev CLI

A terminal-native workflow that feels like kubectl and docker. Deploy, inspect, and evaluate employees without leaving your shell.

$ zev

Zev Daemon

A local control plane that schedules employees, reconciles desired state, and keeps your workforce running on your own machine.

control plane

Employee Registry

Store, version, and share employee manifests. Pull a proven employee the way you pull a container image.

versioned

Runtime Adapters

Run the same employee on Hermes, the OpenAI Agent SDK, LangGraph, or CrewAI. The runtime is a pluggable detail, never a lock-in.

runtime-neutral

Skill System

Package capabilities into reusable, testable skills. Compose them onto any employee instead of hand-tuning fragile prompts.

skills > prompts

Tool Gateway

A broker for credentialed tool access. Employees request capabilities; the gateway enforces scopes, secrets, and audit — by design.

governed access

Memory System

Short-term, working, long-term, and experience memory — so employees recall context, learn from outcomes, and get better with tenure.

4 tiers

Evaluation Framework

Score an employee against its role before it ships. Answers the only question that matters: can this employee be trusted for this job?

evaluation first

// Lifecycle

Manage employees like people, ship them like code.

Every digital employee moves through a governed lifecycle — from the first hire to a graceful retirement — with evaluation gating every promotion.

01
Hire
Author the manifest and onboard the employee into your registry.
02
Train
Attach skills and memory; shape behavior with real tasks.
03
Deploy
Roll it out on the daemon and put it to work.
04
Evaluate
Score performance against the role before trusting it.
05
Upgrade
Swap runtimes, add skills, and re-evaluate safely.
06
Transfer
Move a proven employee across teams or environments.
07
Retire
Decommission gracefully and archive its record.

// Flagship employee

Meet ZevOps, your AI SRE.

The first flagship employee ships in the box. ZevOps watches your systems and works an incident end to end — so a page at 3am gets a head start before a human even opens their laptop.

Triages alerts and correlates related signals
Inspects logs and queries metrics for context
Reasons toward a probable root cause
Files a GitHub issue with the evidence attached
Drafts a clear incident report for the on-call human

Deploy ZevOps →

zevops — incident #4821

# alert received · api-gateway p99 latency
zevops › triaging alert ALT-4821 …
zevops › querying grafana metrics (last 30m)
zevops › inspecting kubernetes logs for svc/api-gateway
  → connection pool exhausted after deploy v2.14.0
zevops › root cause: DB max_connections not scaled with replicas
zevops › filed github issue #4821 ✓
zevops › drafted incident report ✓
# awaiting human approval: production-change

// Quick start

From zero to a running employee in three commands.

Install the CLI, deploy the bundled example, and evaluate it. If you know kubectl, you already know ZevZev.

bash — ~/zevzev

$ zev deploy examples/zevops
employee "zevops" deployed ✓

$ zev get employees
NAME     ROLE          STATUS
zevops   SRE Engineer  Running

$ zev eval zevops
Evaluation: zevops
Score: 82/100
Result: Passed

zev is local-first — everything above runs on your own machine, no account required.

// Principles

Opinionated where it counts.

Five design principles shape every decision in ZevZev — and every employee you build on it.

Local First

Your workforce runs on your machine and your infrastructure. No mandatory cloud, no lock-in.

Runtime Neutral

Hermes, OpenAI Agent SDK, LangGraph, CrewAI — the runtime is a swappable adapter, not a dependency.

Evaluation First

Nothing is trusted until it is scored. Evaluation gates every deploy and every promotion.

Skills Over Prompts

Capabilities are packaged, tested, and reused — not buried in brittle, copy-pasted prompt strings.

Governance By Design

Permissions, approval gates, and audited tool access are built in from the first line of the manifest.