You can't see it
Which files an agent read, which commands it ran and where it connected are rarely shown to you.
Our Terms of Use change on 1 December 2026, and our Privacy Policy and Subscription Agreement on 15 and 21 December 2026. The updates cover account holders, fraud and abuse prevention, including device checks at signup and login, optional sign-in text messages, Hosted Inference for Fleet organizations, how Kenaz connects to AI tools and platforms made by other companies, automatic memory capture, personal files, encryption, and what your organization’s administrators and team leads can see. Privacy Policy · Terms of Use · Subscription Agreement · About text messages · Questions: dpo@kameas.ai
FreeNo accountmacOS on Apple Silicon
Kenaz gives an AI agent its own virtual machine on your Mac, not the run of your machine. A profile sets what it can reach, policy checks what it does, and a signed ledger on your Mac records what happened.
Not on a Mac? The open-source Kenaz Harness runs on its own on macOS, Windows and Linux.
Agents and their tools run in a Linux virtual machine, with one folder you choose shared in.
Allowlist or air-gap a workbench. The VM's firewall drops everything else and Kenaz shows you what it blocked.
Every tool call is checked against Cedar policy before it runs, with five autonomy levels on top.
Workbench events and policy denials land in a hash-chained, Ed25519-signed ledger on your Mac.
Why a workbench
Which files an agent read, which commands it ran and where it connected are rarely shown to you.
An agent on your laptop runs with your credentials, your files and an open network.
When something goes wrong, there is no record of what ran, which files it touched, or where it connected.
The workbench
A workbench is a Linux virtual machine that Kenaz starts on your Mac from a profile and opens in its own window, with one tab per app. The agent works in there. The Kenaz desktop app runs on macOS 13 or later on Apple Silicon, and is free with no account.
01
A profile is a saved recipe: which apps run inside, how much CPU, memory and disk the VM gets, which folder on your Mac is shared into it, which AI provider the Harness uses, and which network destinations it may reach.
Egress tier
The VM's local-only policy. While you are signed in to Kenaz it is treated as allowlist, so the workbench can reach your sign-in server and the destinations this profile allows.
The VM reaches only the destinations allowed for this profile: endpoints you add, the hosts its apps and MCP connectors need, and an optional preset.
All outbound traffic from the VM is blocked. Connectors are turned off and none of their credentials are passed in.
Example profile. Rules are set in the VM's firewall when the workbench starts.
02
The Harness checks every tool call against Cedar policy first. Inside a workbench the VM is the boundary, so the agent has a shell by default and asks before a command no rule covers. An organization's Fleet policy can turn the shell off, and takes precedence over local defaults.
File reads and edits run without asking. Shell commands, network calls and destructive actions ask first.
// No MCP servers started through npx. forbid ( principal == User::"local", action in [Action::"add_recipe", Action::"spawn_recipe"], resource is MCPRecipe ) when { context.recipe_command == "npx" };
A forbid rule wins at every autonomy level.
03
The host keeps an append-only ledger of workbench events and policy denials. Each row is chained to the one before by its SHA-256 hash and signed with an Ed25519 key kept on your machine, so editing or deleting history breaks verification.
The Harness keeps its own hash-chained audit log of every prompt, response, tool call and permission decision. You can verify the chain and export it as JSONL, CSV or PDF.
Example rows.
04
The desktop app runs an observer daemon, sigild, that records activity metadata on your Mac so you can see what happened in a session. It stores it locally and makes no network requests in its default configuration.
| Source | What is recorded |
|---|---|
| Files | paths, sizes, times |
| Processes | what ran, and for how long |
| Terminal & git | commands, repositories, branches |
| Clipboard | type and size, never contents |
| Typing | keystroke counts, never keys |
| App context | frontmost app and window |
Kenaz Harness
The Kenaz Harness is the agent runtime inside every workbench, and a standalone app you can run on its own. It is Apache 2.0 at github.com/kameas-ai/kenaz-harness. The desktop app and Fleet are closed source today.
Models
Bring your own key for any of these, or run local models through Ollama. Keys stay in your OS keychain.
Tools
Built-in tools cover files, shell and web. Connect MCP servers for the rest; each server gets its own grant, and newly added ones ask every time.
On its own
Run outside a workbench, a fresh install reads files and edits inside the workspace without asking. Shell commands, network calls and writes outside the workspace ask first. You can loosen or tighten it.
Platforms
Signed installers for macOS and Windows. Linux builds are unsigned and best-effort. Free, with no account.
Architecture
The desktop app is the control plane, the host keeps the record, and the workbench VM runs the agent.
launches the VM · applies policy
events over vsock · egress filtered
On Linux there is a desktop build, but Kenaz does not start a virtual machine there: workbenches run as processes on the host.
Privacy
On your machine by default
What can leave, and when
Read the code
The Kenaz Harness, which runs your agents, is open source at github.com/kameas-ai/kenaz-harness. What Fleet receives is listed field by field in the Telemetry Disclosure.
For teams · Kameas Fleet
Fleet is the optional, paid service behind Kenaz. Everything on your machine keeps working without it.
Push workbench policy and Cedar rules to every device, see which devices applied it, and lock down in an emergency.
Adoption tiers, session counts, tool mix, as team totals or per person, where your organization has it on.
Model usage and cost by team, or per person where your organization has it on.
An audit log of administrator actions in Fleet, next to each device's signed workbench ledger.
No account
The desktop app and the Harness, fully usable on your own machine.
DownloadIndividual
Hosted inference, sync of settings, context and memory, and a personal Fleet dashboard.
See pricingTeams
Shared team context and one workbench policy across every seat.
See pricingEnterprise
Custom Cedar policies, organization-wide context, and SSO (Google, Microsoft, SAML).
Talk to usKameas hosts Fleet. Self-hosted and dedicated deployment is on our roadmap. Compare every plan.
Questions
No. The desktop app and the Harness are free and work without one. Kenaz offers a sign-in on first launch; you can skip it and sign in later. An account matters only for Kameas Fleet, which is paid.
The desktop app needs macOS 13 or later on Apple Silicon. There are no Windows or Intel Mac builds of the desktop app. The Kenaz Harness runs on its own on macOS and Windows; its Linux builds are unsigned and best-effort.
On a Mac, workbench VMs run on Apple's Virtualization framework through Tart. Kenaz offers to install Tart the first time you open a workbench. Tart has its own licence; see the first-run guide.
Yes. On the air-gapped tier all outbound traffic from the VM is blocked. On the allowlist tier it reaches only the destinations the profile allows. One thing to know: web pages shown from workbench apps load in the desktop window, outside the VM's firewall.
Kenaz does not install updates itself. It checks the release list on downloads.kameas.ai at most once a day and tells you when a newer version is out. The check sends no account or device ID, and you can turn it off.
Pro and Team start with a 14-day free trial. We collect a payment method at signup and don't charge it during the trial; cancel before it ends and you pay nothing. Full terms are on the pricing page.
The desktop app and the Harness are free with no account. Pro and Team add Kameas Fleet and start with a 14-day free trial: we collect a payment method at signup and don't charge it during the trial; cancel before it ends and you pay nothing. Need a custom enterprise rollout instead? Talk to us.