Unsafe Tool Auto-Approval Preferences

Tools that AI Positron cannot undo or preview after they run are considered unsafe. These tools include all tools obtained through the Model Context Protocol integration, tools that run command lines, and tools contributed by other plugins and add-ons (see Tools Contributed by Plugins). By default, the end user is explicitly asked in the chat box to allow these tools to run. When prompted, the end user can also choose to always allow a tool to run. In that case, the tool is automatically added to the list of allowed tools on this preferences page. The search_chat_history tool, which searches your past conversations, is handled the same way because those conversations may contain content that you did not intend to share again.

The text area contains a list of tools that are always allowed to run without asking the end user, with one tool per line. Each entry begins with a tool name, which must match exactly unless it contains the * character. The * character is a wildcard that matches any sequence of characters.

For the command-line tool (run_command_line), the tool name can be followed by a space and a command key. A command key is a short, underscore-separated name that the AI model assigns to each command line. It describes the command line rather than repeating the command line. For example, the command line p4 opened //demo/spec/... may be assigned the p4_opened_demo_spec key. When an end user chooses to always allow a command line, an entry containing the tool name and command key, such as run_command_line p4_opened_demo_spec, is added to the list. This lets the end user always allow a specific command line while still being prompted about other command lines. The command key in an entry is matched as follows:

  • Without a wildcard, the entry's command key must be identical to the command-line key. For example, run_command_line p4_opened allows the p4_opened key but not the p4_opened_demo_spec key.
  • With a wildcard, the entry's command key is a pattern that must match the entire command-line key. For example, run_command_line p4_opened_* allows all keys that start with p4_opened_, and run_command_line p4_* allows all keys that start with p4_.
  • Without a command key, the entry allows all command lines. This applies to both run_command_line and run_command_line*.

The AI model chooses the command key to describe the command line. The key does not restrict the command line itself. Because the AI model chooses the key, the same command line may receive a slightly different key in a later run. Use a wildcard to cover these variations. The Always allow option is unavailable for a command line whose key contains the * character.

For MCP tools, each tool call must initially be approved individually. For example, when this tool call is always allowed:



the preferences page is updated to contain this entry:



If you want to allow every tool call to the MCP server, you can manually edit the entry to use the wildcard form brave-search_*.

An entry consisting only of * characters, or a command key consisting only of * characters, matches nothing. Therefore, it cannot accidentally allow every tool or command line.

You can save this preferences page at the project level to share a specific set of allowed tools with team members.

Tools Contributed by Plugins

Before the AI runs a tool contributed by another plugin or add-on, the chat asks for your confirmation, even when the plugin declares that the call is safe. For such a call, the approval box asks Allow this tool to run?, shows the name of the tool and that it comes from a plugin, describes what the tool does, and shows the arguments of the call, if any. The tools that Oxygen XML Web Author contributes, such as compare_documents, are not affected.

For a call that the plugin declares safe, the approval box offers the following choices:

Always allow safe calls of this tool
Runs the tool and adds it to the list on this preferences page, so that calls the tool declares safe no longer require confirmation. The calls that it declares unsafe still ask for your confirmation. Pressing Enter has the same effect.
Allow once
Available in the menu of Always allow safe calls of this tool. Runs the tool only this time. When the call cannot be always allowed, it is the only choice besides Skip, and pressing Enter has the same effect.
Always allow safe calls to tools in the [category] category
Available in the menu of Always allow safe calls of this tool when all the tools of the plugin are in a category of their own that no other tool uses. The menu item names the category, for example Always allow safe calls to tools in the Git category. Runs the tool and allows all calls that the tools in the category declare safe. Calls that they declare unsafe still require your confirmation. This also covers the tools added to the category later by an update of the plugin, or by another plugin that takes its place. While two plugins contribute tools in the same category, it applies to neither of them.
Skip
Does not run the tool, and tells the AI that the call was skipped. Pressing Esc has the same effect.

A call that the plugin declares unsafe requires approval in the same way as calls from other unsafe tools: the approval box shows the plugin's own warning (or, if the plugin gives none, the name of the tool and the arguments of the call) and offers Allow, which runs the tool this time only, and Skip. Pressing Enter has the same effect as Allow, and pressing Esc has the same effect as Skip. The menu of Allow offers to always allow the tool, which adds an entry that allows every call of the tool, such as plugin:git_push. The calls that a plugin declares unsafe can only be always allowed for each tool, never for a whole category.

The tools contributed by plugins are listed on this preferences page under distinct names, so that an entry added for another tool never allows them:

  • plugin: followed by the tool name, such as plugin:git_push, allows every call of the tool.
  • The same entry followed by a space and safe-calls, such as plugin:git_status safe-calls, allows only the calls that the tool declares safe.
  • plugin:category: followed by the category, in lower case and with underscores in place of its spaces, such as plugin:category:git, allows the calls that the tools of the category declare safe.

An entry containing the * character applies to tools contributed by plugins only when it starts with plugin:. For example, plugin:* allows all of them, while *_status does not allow plugin:git_status.

When the user cannot be prompted for confirmation (for example, in a transformation that uses the AI XPath functions or in an action that the AI runs by itself), a tool contributed by a plugin runs only if it is always allowed, even for the calls that the plugin declares safe. A call that the plugin declares safe is allowed by an entry such as plugin:git_status safe-calls or by the category of the tool, while a call that it declares unsafe is allowed only by an entry such as plugin:git_push. Otherwise, the AI receives a refusal that identifies the entry to add and logs the refusal.

To stop being prompted for confirmation before tools contributed by plugins run, clear the Ask for confirmation before running the tools contributed by plugins option in the Installation and Configuration preferences page.