Skip to content

AgentZ¶

AgentZ logo AgentZ Say it "agent zee"

A zero trust runtime for AI agents.

Every agent runs in a default deny sandbox, holds no secret, and writes every action to a trace you can replay. Open source, and free to start.

Start free Star the repo Watch the demo AgentZ docs Product page

AgentZ is the AccuKnox platform for production AI agents. Describe a job in a sentence, and AgentZ writes the skill and wires the steps. The agent then runs from chat, an API, the CLI, or a cron.

Zero trust means the agent gets no access until a policy grants it. Default deny means every network call and tool call is blocked until you allow it. The policy runs at the kernel, next to the agent, so nothing leaves the sandbox without a record.

Full product documentation lives at docs.agentzharness.ai.

Four steps¶

01

Build

Describe the job. AgentZ writes the skill and wires the steps.

02

Run

Chat, API, or CLI. Any framework, and no redeploy.

03

Automate

Trigger on a cron, an event, or an API call. Skills chain.

04

Govern

The kernel checks every call. The trace records the result.

One control plane¶

Skills

Reusable, versioned building blocks.

Workflows

Chain steps, schedule them, hand off work.

Context

Shared memory, files, and knowledge.

Teams

Roles, ownership, and a shared scope.

Guardrails

No standing access to any tool or credential.

Audit

Every step recorded, span by span, and replayable.

The zero trust part¶

Sandboxed on run one

Isolation is not a setting you turn on later.

No secret in the agent

Credentials are scoped and injected at run time. An injection cannot leak what the agent never sees.

Egress control at the kernel

Allow or block by domain, port, and protocol.

Permission per action

Read and scan pass. Mutate, push, and delete are denied by default.

Roles gate tool calls

Fine grained RBAC covers each agent action.

On-premises and air-gapped

Logs and audit evidence stay on your infrastructure.

Nine screens¶

Model picker listing GLM, Claude, Gemini, Kimi and GPT models available to an agent
Sandbox update screen with per-tool toggles for each connected MCP server
MCP connection form with Slack, GitHub, Notion, Linear, Asana, Figma and Atlassian servers
A cloud asset count diff report generated by an agent, next to its output files
Schedule editor with a cron expression, timeout and run history limits
Workflow run graph with a step inspector showing status, timings and instructions
Span list with model calls, bash and webfetch, and a token breakdown for one call
Graph of MCP tools called by an agent, with latency and last-used age per tool
AgentZ closing slide with the SOC 2, CNCF, AWS and Nutanix certifications

Your stack, your model¶

Credentials, scopes, and policy live on AgentZ instead of on the agent. Run a frontier API or a self-hosted open weight model on your own key, and switch models without rewiring anything.

Connectors ship for Slack, Gmail, Microsoft 365, Google Workspace, Jira, Confluence, Notion, GitHub, GitLab, and Bitbucket. Any MCP server works after you authorize it once.