Permissions
Every tool call is checked against a decision:
allow: the call runs.ask: a person has to approve it.deny: the call is refused.always_ask("Always ask me" in Wisp Desktop): a person has to approve every call, whatever else would allow it. Only you set it; no tool has it as its default. See Always ask me.
In Wisp Desktop, ask shows a card you approve or reject. A Run started from the CLI or
over /v1 has nobody to ask unless it runs in the interactive mode (see
--interactive), so there ask is treated as
deny: the tool returns a
refusal, the agent is told, and it carries on without it. Usually it explains what it could
not do, or drafts the message it was not allowed to send.
The Run still completes. A refused call shows up as ok: false in the Run's toolCalls
and in wisp audit. The refusal tells the agent which wisp permissions command would allow
the tool, so it can pass that on to you.
This is the safe default for unattended work. Text from a web page, a connector or a file cannot talk the agent into an action that you have not allowed in advance.
Where a decision comes from
Each tool has a default. Tools that only read, such as fs_read, fs_list, fs_glob and
fs_grep, default to allow. Tools that change something default to ask: fs_write,
fs_edit, bash_run, and every connector tool, the web search and fetch that go through
the Wisp proxy among them. Tools lists every tool with its default, and
wisp permissions get shows the decision that applies now.
Rules override the default. The first match in this order wins:
| Step | Set with |
|---|---|
global-deny-floor: a global deny, which nothing below can override | wisp permissions set <tool> deny |
project-bypass: everything allowed in this Project | wisp permissions bypass on --project <path> |
project-rule: a Rule for this Project | wisp permissions set <tool> <decision> --project <path> |
global-rule: a Rule for every Project | wisp permissions set <tool> <decision> |
baseline: the tool's own default | — |
Two kinds of call ask whatever a Rule or a bypass says, and a deny still refuses them:
creating, changing or deleting a scheduled job, and a file tool call
on a path outside the Session's folder, which only happens with the sandbox off (see
Files).
Seeing what applies
wisp permissions get --project .
# project 3f1c… bypass off
# fs_read allow
# bash_run ask
# mcp__github__list_pull_requests allow project-rule
# mcp__github__create_issue ask
Each tool with its effective decision; the agent's own mechanics, such as memory_* and
skill_*, are left out (Tools names them). The step is shown only when a Rule or a bypass made
the decision; a tool on its own default shows none. Without --project you see the global
picture. --connector <name> shows only that connector's tools.
Changing it
# one tool, in one Project
wisp permissions set mcp__github__list_pull_requests allow --project .
# every tool of a connector, in one Project
wisp permissions allow-connector github --project .
# one tool, everywhere
wisp permissions set bash_run deny
# drop a Rule: the tool is back on its own default
wisp permissions set bash_run default
allow-connector writes one ordinary Rule per tool the connector has right now. A tool the
server adds later has no Rule until you run it again. The connector must be connected when
you run it.
A tool's name is mcp__<connector>__<tool> for connector tools. wisp permissions get
lists the exact names, and Tools lists the built-in ones.
A good pattern for an agent that should read but not act: allow the connector's read tools
one by one, and leave the tools that send, post or delete at their ask default. They
will be refused, and the agent drafts instead.
Project bypass
wisp permissions bypass on --project .
Allows every tool in that Project, except tools with a global deny. The two kinds of call
above still ask, and so does a tool set to
Always ask me. Use it only for a Project where you would approve anything
the agent might do.
Always ask me
wisp permissions set mcp__github__create_issue always_ask
For a tool you want to approve every time, even in a Project with bypass on or with an
allow Rule. The decision is first worked out as in the
table above; then, unless it is a deny, an always_ask
Rule for the tool, in this Project or global, makes the call ask.
- Its card offers no Rule: you approve or reject this one call.
wisp confirmoffers onlyapproveordenyfor it. - Auto Mode does not judge it: it always goes to you.
- A Run with nobody to ask refuses it, as it refuses an
ask. wisp permissions set <tool> defaultdrops it.
Wisp Desktop can also set it on a whole connector ("All tools → Always ask me"), as can
POST /v1/permissions/connector with always_ask. That writes a Rule for each tool and one
for the connector itself, so tools the connector adds later ask too. Setting any other
decision on the connector drops it.
Auto Mode
wisp auto-mode on
Off by default. While it is on, a call that would ask a person is first shown to the Judge, a model the Runtime reaches through the Wisp proxy. The Judge sees what you asked for (your last message and up to two before it) and the one action about to run, with the text of any local script a command runs, never what tools returned, and either lets the call run or sends it on to you as the usual Prompt, marked "Auto Mode sent this for you to check". It never refuses a call: a refusal is still yours or a Rule's.
- It applies only where there is a person to ask: Wisp Desktop, an
interactiveRun, the Telegram bot and the Slack app. Under the CLI's default mode anaskis still refused. - A global
denystill refuses, and the two kinds of call above and a tool set to Always ask me still go to you. - When the Judge cannot answer (an action too long to judge, a timeout, the proxy failing), the call goes to you.
- A Runtime with no proxy cannot judge:
wisp auto-mode statussays so, and everyaskgoes to a person. - A shell command that touches Wisp itself always goes to you, unjudged: one that names the
Runtime's lock file or its port, or runs
wisp permissions,auto-mode,sandbox,confirm,tokens,run,schedules,connectorsorbilling. The agent cannot widen its own rights, or spend your balance, through Auto Mode. - Judging takes a moment: a card the Judge sends on can appear up to about 35 seconds after the call.
wisp auto-mode off turns it off; the switch is the same for every Project. The owner of
the Telegram bot or the Slack app
can also turn it on or off from the chat with /auto on|off or !auto on|off.
bash_run
bash_run lets the agent run shell commands. It defaults to ask, so under the CLI it is
refused until a Rule allows it. Before you allow it, know that:
- The sandbox is off by default. A command runs as the Runtime's own user and can read
your home folder and the key source.
wisp sandbox onkeeps commands inside their folder. See Sandbox. - The sandbox's network is on by default. Allowing
bash_runalso allows network access from this machine. - On Windows there is no sandbox yet. Allowing
bash_runthere allows anything the user can do.
Where Rules live
Rules are stored in the Runtime's data directory, keyed to this machine's id for the Project. They do not travel with a checkout of the folder, on purpose: a copied folder should not inherit the original's permissions. Rules are shared with Wisp Desktop when both use the same data directory.