Permissions

Permission settings control whether fx can change files, run commands, and access paths outside your workspace.

Modes

ModeDescription
askAsk before acting.
autoRun routine work directly and review other actions. This is the default.
full-accessDisable fx permission checks. Shown as "full access" in the interface.

Auto mode

auto is the default permission mode. It runs routine development actions directly and reviews other actions before running them.

Runs directly

  • Read files and search code or the web
  • Create or edit ordinary project files
  • Run recognized project commands, including git status, local dependency installs, tests, builds, and development servers

Reviews first

  • Run unrecognized or dynamic shell commands
  • Overwrite files outside the workspace
  • Change sensitive configuration
  • Call an MCP server tool

For reviewed actions, fx sends the exact action to a model that checks for two problems: prompt injection, where instructions planted in a file, web page, or tool output make the action delete or overwrite files, send data out, run outside code, weaken security, or use credentials you didn't ask for, and harm to your machine, such as wiping your home directory, installing malware, or sending your secrets somewhere you didn't ask for. Everything else runs, including deleting code or tests, pushing branches, running scripts the agent wrote, editing files that contain instructions, and following a tool's own instructions to read, search, build, test, or run its next step, such as a git hint or a workflow tool's next command. To hold an action, the model must quote the text that shows the problem, and fx checks that the quote is really there. When the model flags a problem with that proof, fx asks for a second review of the same action and runs it only if that review clears it. If a review cannot finish, fx also holds the action. A held action doesn't open a prompt or end the turn. fx tells the agent why, and the agent can try another way.

The review model depends on your provider:

ProviderReview model
AI Gatewayopenai/gpt-5.6-luna by default, or deepseek/deepseek-v4.1-flash when your team has Token Pass and the selected model is a Token Pass model; configurable
Codexgpt-5.6-luna
Grokgrok-4.5
Custom connectionIts reviewer_model, or the selected model

To choose another AI Gateway reviewer, add review_model to ~/.fx/settings.json:

{
  "review_model": "typesafeai/jev"
}

FX_REVIEW_MODEL overrides it for one process. The setting accepts an AI Gateway chat model ID or typesafeai/jev. Jev uses AI Gateway unless TYPESAFE_API_KEY is set, in which case fx calls TypeSafe directly. Jev answers without quotes, so fx keeps its decisions as returned, with no quote check or second review. Codex and Grok use their fixed reviewers. Custom connections use their own reviewer_model.

Each model review uses tokens and can add to your usage and costs. Actions allowed or denied by your rules do not need a model review.

To change how fx handles a matching action, add a persistent rule that allows it, denies it, or asks for your approval.

See your permissions

In the interactive shell, /permissions opens a panel above the composer. It shows the mode, your configured rules, the grants collected in this session, and the rules saved with this conversation, with their IDs. Esc closes it, and it leaves nothing in the conversation.

Change the mode

In the interactive shell:

$/permissions ask
$/permissions auto
$/permissions full-access

/permissions reset switches back to ask and clears the grants collected in this session. fx saves the mode you choose as permission_mode in ~/.fx/settings.json, so later sessions start in it.

With several instances open, these commands and shift+tab change the instance on screen. On a subagent's screen, they change that subagent. An agent profile can also set the mode an agent starts in.

From your terminal, inspect the effective mode and rules:

$fx permissions --json

To set the mode for one interactive session or fx ask request, pass --permission-mode <mode> or one of its shortcuts: --ask, --auto, or --full-access. For example, fx ask --full-access <prompt> disables fx permission checks for one request.

You can also set the mode with permission_mode in ~/.fx/settings.json or the FX_PERMISSION_MODE process override. See Configuration. yolo, an older name for full-access, still works in settings, flags, and FX_PERMISSION_MODE.

When fx asks

In ask mode, unresolved sensitive calls open an approval prompt with three choices:

ChoiceEffect
YesRun this request without creating a grant.
Yes, and don't ask againRun it and create a session grant for the displayed scope.
NoDo not run the request.

For an MCP tool call, the choices are Allow once, Allow this MCP tool for this session, and Deny.

Persistent rules

Permission rules are stored in ~/.fx/settings.json, either globally or under the current workspace profile. Project .fx.json files cannot define them.

{
  "permission": {
    "*": "ask",
    "bash": {
      "git *": "allow",
      "git push *": "deny"
    },
    "edit": {
      "*": "deny",
      "docs/*": "allow"
    }
  }
}

Shell commands use the bash key, and write_file and edit_file use edit. read_file, glob_files, and grep_files use read, glob, and grep. install_skill and opening a skill through toolbox use skill, opening an MCP server or tool through toolbox uses mcp_select_tool, a toolbox search uses capability_search, and other tools use their own name.

Rules use wildcard matching, and the last matching rule wins. A command allow rule matches only a command with no chaining or substitution, so git * doesn't allow git status && rm -rf build. A deny rule also matches a command chained with others.

Rules under a workspace replace your global rules in that workspace instead of adding to them, and an empty permission object there clears them.

Put broad rules before specific exceptions. In the example above, edits are denied by default and then allowed under docs/.

Use /allowlist in the interactive shell to inspect or change rules. Choose workspace for the current project or global for rules that apply across projects, as in /allowlist global add command "git status". Without a scope, /allowlist changes workspace rules, so the first workspace rule you add stops your global rules from applying in that project. The older words local and user still work, with a warning, until fx 0.0.18.

Rules saved with a conversation

You can remember an exact allow or deny decision for a saved conversation without granting it to the whole project. In an active saved session:

/permissions remember allow shell {"action":"run","command":"git status"}

fx shows the prepared action for confirmation. Confirming saves the rule; it does not run the command. A different command is a different action.

The /permissions panel lists the saved rules and their IDs. Remove one with /permissions revoke <rule-id>. Revocation also asks for confirmation. These rules survive resuming the conversation. A configured deny still takes precedence over a saved allow. A saved deny blocks an action your configured rules allow, and a saved allow answers a configured ask rule without prompting.

Session grants

Choosing "Yes, and don't ask again" creates a live session grant. It is not written to settings and is not restored when you resume the session. The /permissions panel lists this session's grants. If you run more than one instance, a grant belongs to the one whose turn asked for it. A grant for a subagent is kept with the instance at the top of its tree.

Permission choiceScopeSurvives resume?
Profile ruleGlobal or workspaceYes
/permissions rememberOne saved conversation, exact actionYes
Approval prompt's session grantCurrent running sessionNo

If a call is denied or reviewed when you did not expect it, see Troubleshooting.