Permissions
Every tool call goes through a permission check before it runs. You can pre-grant categories or paths to cut down on prompts without giving up auditability.
Defaults on a fresh install
What the Harness does before you change anything depends on where it runs.
On your computer (the Harness on its own):
- The shell tool (
bash) is off until you turn it on in the Tools view. When on, it runs only commands on its allowlist. - Web search and web fetch are off until you turn them on.
- The Default autonomy level runs file reads and edits without asking. It asks before shell commands, network calls and destructive actions.
- The Cedar policy layer allows everything unless you, or your org through Fleet, add
forbidrules. Settings → Policy shows the active rules.
:::note Launch default A starter policy for the Harness on your computer has been decided and is being built. It allows reads and edits inside your workspace, and asks before shell commands, network calls and writes outside the workspace. When it ships, existing installs keep their current behaviour until you accept the new default. :::
Inside a Kenaz workbench:
- The virtual machine is the boundary. The Harness can reach only what the workbench's network rules and shared folder allow.
- The shell tool is on by default. A shell command that no rule permits or forbids asks you first.
- An org admin can turn the shell tool off for the org with a Fleet policy
forbidrule. The change applies on the next policy poll.
On either, a policy pushed from Fleet takes precedence over the local default.
The categories
| Category | Covers |
|---|---|
| Filesystem read | read_file, list_dir, path_complete, MCP filesystem reads |
| Filesystem write | write_file, MCP filesystem writes |
| Shell | bash and any MCP server that exposes a command-execution tool |
| Network | web_search, web_fetch, MCP servers that talk to external HTTP |
| Specific MCP servers | Per-server grants (e.g. "GitHub: enabled", "Slack: ask each time") |
A tool call must clear both the category and any path-specific rules.
Granularity
Three levels, from most to least restrictive:
- Ask each time — modal pops up showing the exact arguments. Approve / deny / approve-all-for-session.
- Approved for session — the tool runs without prompting until the session ends, then resets.
- Always allowed (with scope) — never prompts. Most useful for filesystem reads scoped to a project root.
You can also explicitly Deny a category — Kenaz will not prompt and the tool will fail closed.
Scoping filesystem grants
For Filesystem read/write, grants can be scoped:
- Anywhere — risky, only do this knowingly.
- Within these paths — e.g.
/Users/you/code/project-aplus/tmp/project-a-work. The check is a strict prefix match after symlink resolution. - Project root — the path attached to the active project. Switches automatically when you switch projects.
A tool call against a path outside the granted scope falls back to "Ask each time" — you can approve the one-off without widening the scope.
The dangerous-path tier
Some paths are flagged as dangerous and cannot receive a permanent "Allow always" grant. The list covers common credential and system directories:
~/.ssh/ ~/.aws/ ~/.config/ ~/.gnupg/
~/.netrc ~/.gitconfig
/etc/ /private/etc/ /private/var/
/System/ /Library/ /usr/
/bin/ /sbin/
/proc/ /sys/
When a tool call targets one of these paths the permission modal shows a hazard banner explaining the risk. You can still approve the single call — the tier only blocks permanent blanket grants, not one-off approvals.
Per-MCP grants
Each MCP server gets its own grant configuration. Open the Tools view and expand the server entry. For a server like notion-mcp:
- Always ask (default for newly-added)
- Approved for session
- Trusted — never prompts; tools fire silently. Use only for servers you wrote or audited.
Safe-by-default writes
Filesystem write grants come with two extra guards Kenaz applies independently of the model:
- No overwrite without confirmation — even with a path-scoped write grant, an existing file triggers an "overwrite?" check. You can pre-grant overwrite per-file, but the default is the safe one.
- Atomic writes — Kenaz writes to a
.tmpsibling and atomically renames. If the tool crashes mid-write, the original file is intact.
Audit trail
Every permission decision lands in the audit log — the prompt, the arguments, what you decided, and the timestamp. So even when you've pre-granted a category, the record of every individual call is still there.
Open the Audit view → filter kind = "permission.decided" to review.
Suggested starter setup
For most users, after the first day:
- Filesystem read — Always allowed → Project root only. The model can read code files without pinging you.
- Filesystem write — Always allowed → Project root only. Same idea.
- Shell — Ask each time. This is the dangerous one; don't pre-grant.
- Network — Approved for session. Web search is rarely problematic.
- All MCPs — Always ask until you've used them enough to know what they call. Then promote individually.
You can change any of this any time from the Permissions view; nothing is sticky beyond what you set.