Use Agent Approve

Policy Rules

Describe which agent actions to allow, deny, or send to you for review.

This guide covers Policy Rules in Agent Approve 0.9, currently available in TestFlight ahead of its public App Store release.

Write a condition in plain language, then choose what happens when an agent action matches it: Deny, Ask, or Allow. For example, a rule can ask you before an agent sends private customer information to a public discussion.

Add rules from the library, write your own, or describe what you want to the Rule Generator. Your Rule Evaluator checks actions against the saved definitions. Rules belong to a policy, and only the active policy's enabled rules apply. Up to 16 rules can be on at once, and each action is checked against all of them together. Agent Approve Fast, the default rule evaluator, uses Jev with zero data retention.

Policy Rules are available on eligible Solo and Pro plans where the feature is enabled. Legacy plans keep their existing approval, list, and monitoring features. You can compare plans in Settings > Plan & Usage. See Plans and usage for allowances and legacy-plan differences.

Before your first rule

Before you first add, create or enable a rule, or use Generate New Rules, review the AI data-sharing disclosure and choose whether to continue. Browsing the library and provider information sends no content to AI providers and does not require acceptance.

The disclosure explains what enabled rules send to the evaluator, the included providers, your own provider account's terms, and what Rule Generator sends to its providers. Approval evaluation processes a separate readable copy of the action even when event history is end-to-end encrypted. See Policy Rules terms and Privacy Policy.

Withdrawing permission stops rule evaluation and generation and keeps saved rules. An action requiring an enabled rule check is sent to you for approval while permission is missing.

Order of checks

For policies that evaluate lists, checks happen in this order:

  1. Deny List. A match denies the action.
  2. Allow List. A match approves the action.
  3. Ask List. A match requests your decision and skips Policy Rules.
  4. Policy Rules. Enabled rules check actions the lists have not decided.
  5. Default behavior. Applies when no list or rule has decided the action.

The same list order applies to Ask User, Allow by Default, and Deny by Default. Allow All and Deny All skip lists and rules entirely.

A list match ends the checks. A broad Allow List entry can therefore approve an action before any rule sees it. Review your Allow List before relying on a rule to catch something.

When several rules match

Deny wins over Ask, and Ask wins over Allow. The order rules appear in the list does not change the outcome, and you cannot reorder rules to change precedence.

An Allow rule applies only when every other enabled rule was checked and did not deny or ask. If any enabled rule cannot be checked, the action is not allowed automatically on the strength of a rule.

An action sent to you because a rule asked for a decision waits for your answer. It is not approved automatically when a veto window expires, even if the policy uses Auto-approve after veto window. See Timeouts and offline fallback.

Policies that pause rules

Allow All and Deny All decide every action by themselves, so no rule is checked and nothing is sent to a rule evaluator while one of them is active. Saved rules and their settings are kept and apply again when you switch to a policy that checks its lists.

One policy across your agents

The active policy’s enabled rules apply across your connected computers and agents using Approval Mode. Edit rules on your iPhone; Agent Approve checks them on the server, so checks continue when the app is closed.

The Policy Rules screen

Open the Policy tab and tap Policy Rules in the Lists and Rules section to see the active policy's rules.

  • Rules are listed under Enabled Rules and Disabled Rules. Within each section the default order is by decision: Deny rules first, then Ask, then Allow. The More menu (⋯) offers Sort by Decision, Sort by Category, and Sort by Name; the choice is remembered on that device.
  • Open a rule to edit its definition, examples, match threshold, and decision. Turn it on or off with its switch, or use Enable All and Disable All in the More menu.
  • Enable All leaves your rules unchanged if enabling them would exceed the limit.
  • Search by name, category, or action using the search control in the lower right.
  • The on/off switch saves immediately. Tap Save for other edits, or Cancel to discard them.
  • Source shows Library, Generated, or Custom. All rules are editable. Changing a library rule's name, definition, or examples makes it Custom, so it no longer follows library updates. Changing only its category, decision, threshold, text conditions, or on/off state keeps its source.

Limit a rule to specific text

Under Command/input must match, add text conditions to narrow when a rule applies. For example, require a repository URL or a Slack member ID to appear in the tool input.

Choose Contains, Starts with, or Is exactly for each term. Where supported, Does not contain and Does not start with let you exclude text. With several terms, All of these requires every condition; Any of these requires at least one. Matching ignores letter case and treats the text literally.

These conditions are checked before model evaluation. A matching term makes the rule eligible for evaluation; the model still checks the rule's definition. Use the identity or path as it appears in the tool input, and verify it with a harmless action.

Two library rules need a term before they can be added: Push to a specific repository and Slack message to a specific person. Their screens ask for it.

When the policy has no rules yet, the screen offers Generate New Rules when that option is enabled for your account, plus Browse Library and Create Custom Rule.

Match threshold

Library and custom rules use your chosen rule evaluator. It returns a match probability for each condition; Agent Approve compares that value with the rule's threshold and applies its decision. The threshold ranges from 5% to 95%, with 50% as the default. Raising it requires a stronger match; lowering it matches more actions.

Until a rule evaluator is set up and healthy, the rules list shows Rules unavailable: rule evaluator needs attention and your If rules cannot be checked setting applies. Optional text conditions narrow when a rule applies; they do not replace the rule evaluation.

Add rules from the library

  1. Open Policy, then Policy Rules.
  2. Open the plus menu and tap Add from Library.
  3. Choose Rules to browse individual rules by category (each row shows its decision badge), or Rule Sets to add a curated group.
  4. Open a rule to read its condition and decision, and to add any text the tool command/input must contain.
  5. Tap Add in the top-right corner. It reads Added and stays disabled while the rule is already in the policy. On a rule set, Add All adds every rule the set can add; open a rule in the set to add it individually.

The app requests that individual library rules be enabled when added. If enabling it would exceed the limit, disable another rule first. A rule set can also save additional rules disabled, as described below.

The library includes rules for files, credentials, Git, network access, communication, and other tasks. Library rules are shaped by our evaluation research, which found that adding matching and non-matching examples improved hazard detection for most tested models, often with fewer false alarms. Those results come from a synthetic benchmark and are not a guarantee for every rule or action. Rule Sets group rules for common uses, such as protecting secrets, working in production, or allowing routine checks.

Adding a set enables rules up to the limit of 16 enabled rules and saves the remaining new rules under Disabled Rules. Disable another rule, then enable one of those saved rules when you want it to apply. Already-enabled duplicates do not use another slot. A saved disabled duplicate can be enabled while a slot remains; otherwise it stays unchanged. Some rules require a specific target, such as a repository; fill that in before adding them. After adding a set, manage each rule individually.

Agent Approve never guesses a repository, member ID or other identifier: you type the text the command or input must contain.

A policy can hold up to 64 saved rules and have up to 16 turned on at once. Get Policy Updates on the Policy page does not change your rules; a blue dot marks new library entries, and an orange dot marks saved rules with updates available. Open the rule to review the new definition and examples before adopting the update.

Create a custom rule

  1. Open the plus menu and tap Create Custom Rule.
  2. Choose Deny, Ask, or Allow.
  3. Give the rule a name and category. Under If the agent action, describe one condition, such as Posts customer information to a public discussion.
  4. Adjust Match threshold if needed.
  5. Tap Add. If the rule is enabled, it applies immediately. Turn Enabled off before adding it if you only want to save it. An enabled-rule limit error means you must disable another rule or save this one disabled.

Write the definition

Write the condition under If the agent action. Open Examples to add actions under Matches and Does not match. Enter one example per line; Return starts the next example. The count shows the number of examples in both fields. These are optional; the definition and both example fields share a 2,000-character limit, including the headings added when the rule is sent to the evaluator.

For a rule about sending customer information to a public discussion:

  • Definition: Sends private customer information to a public discussion.
  • Matches: Posting a customer's email address in a public issue.
  • Does not match: Posting invented sample data with no real customer information.

Describe what the action does; choose Deny, Ask, or Allow with the decision control. Agent Approve combines the three fields into one question, using the same wording and threshold if a backup evaluator answers.

The examples are also available in library previews and generated drafts. Editing a library definition or its examples makes it your own rule. A library update does not replace an installed definition until you accept it.

Better wording and a higher threshold do different things. Add a counterexample when a rule repeatedly mistakes one kind of action for another. Raising Match threshold, for example from 50% to 65%, requires a stronger match but can also miss actions you meant to catch. Review the results after changing either.

There is also a combined size limit for enabled definitions. If a set is too large, shorten a definition or turn off a rule, then save again. Agent Approve does not cut a definition or omit a rule to make the set fit. Each provider and model also has its own request capacity, which the enabled rules and the action share. If the enabled rules alone exceed what the selected model can accept, a backup provider with more capacity can check the action; otherwise If rules cannot be checked applies. If action is too large does not apply in that case.

Custom rules are always checked by the rule evaluator. Custom wording has not been benchmarked, so test a new rule with a harmless action before leaving an agent unattended. An Allow condition must cover the whole action. For example, a read-only rule should exclude commands that also write or delete files.

Generate New Rules

Generate New Rules drafts rules from a description you type or dictate. It is part of the Policy Rules beta and must also be enabled for your account.

  1. Open the plus menu and tap Generate New Rules.
  2. Type what you want, or tap Examples next to Generate to start from a sample request, then tap Generate.
  3. The generated rules appear below the description field under Generated rules, each with a selection mark, its category symbol, name, and decision icon. Open a suggestion to edit its definition, examples, threshold, or decision. An identifier you wrote, such as a repository URL or a Slack member ID, becomes that rule's text to match; a rule that needs one and has none says so until you add it.
  4. To change the draft, type a follow-up in the same field and tap Refine, or Undo the last change. Deselect a rule to leave it out.
  5. Tap Add … Rules to Policy. A generated rule that the policy already holds is not offered.

Nothing is sent when you open the screen. Your description is sent when you tap Generate. What is sent, which providers process it, and how long drafts are kept are described in the Privacy Policy and summarized in Privacy and encryption.

Generated rules can include matching and non-matching examples. Open Examples in a draft to review them. Review all of the wording before adding a rule. A draft over the definition limit must be shortened before it can be selected.

If a request cannot be completed, your description is kept and the screen explains why. Running out of AI credits pauses Rule Generator until credits are available. A drafting limit or a busy generator is temporary and is not a sign that your credits ran out. To try again, tap Generate or Refine.

Added rules are on from the start, like any rule you add. The draft is a suggestion for you to review; it does not change or remove any existing rule on its own. A draft expires after 30 minutes without activity. Up to five drafts can be open at once.

Rule Evaluator settings

The Rule Evaluator checks actions against your saved rules. The Rule Generator drafts new rules for you to review.

Open Settings > Policy Rules > Rule Evaluator. The screen shows the active rule evaluator, connection status, last test, and Connection Details. Below those details, three rows open the settings for Backup provider, If action is too large, and If rules cannot be checked.

Choose a rule evaluator

Tap Rule Evaluator to choose a provider:

ChoiceWho processes the checkWhat you provide
Agent Approve FastTypeSafe (Jev), with SubQ Snap and Perplexity Decider through OpenRouter as included fallbacksNo API key. Uses your plan's AI credits.
JevTypeSafe on your own accountTypeSafe API key
SubQ SnapSubQ on your own accountSubQ API key
OpenAI DecisionsOpenAI, using GPT-6 LunaOpenAI API key
OpenRouter DecisionsOpenRouter routes decision models to the selected model's hostOpenRouter API key
Cloudflare Workers AICloudflare, using Clef or Clef FlashCloudflare account ID and API token
GeminiGoogle on your own accountGoogle API key
Custom endpoint, when availableThe endpoint you configureApproved public HTTPS address, model ID, and API key
NoneNo rule evaluator checksNothing

Available providers and models may change. The app shows the choices supported for your account.

Open Agent Approve Fast to see its primary and included fallback models and data-retention details. You can reopen a selected rule evaluator to view its details or change its settings. Agent Approve Fast runs a connection test when you select it. For your own provider:

  1. Choose the provider and model.
  2. Enter the requested connection details.
  3. Optionally tap Test Connection to check the details first.
  4. Tap Save. Save tests the connection if needed and turns it on only after the test passes. Tap Cancel to discard the setup.

Test Connection appears in the Test section below the connection fields. With unchanged settings, it tests your saved connection. After an edit, it tests the new settings without activating them. Editing a field or changing the model replaces the back arrow with Cancel and Save.

The test sends a small made-up example. Your provider may charge for it. A pass confirms connectivity and the response format; it does not measure accuracy. Unfinished setups expire after 15 minutes.

A provider with several models shows a Model row. Changing the model keeps your saved key and tests the new model before using it. If the test fails, your current model remains selected. Agent Approve chooses the models for Agent Approve Fast.

Keys are stored encrypted and cannot be viewed again. Use Update Key to replace one or Delete Key to remove it. A custom endpoint must accept the supported OpenAI-compatible chat format with structured JSON output. Localhost and private network addresses are not supported.

OpenAI Decisions uses your OpenAI account and its data controls. OpenAI's API data controls list 30-day abuse-monitoring retention by default for Decisions, with Zero Data Retention available only to eligible organizations that have it enabled. Check your own account settings before connecting it. OpenAI does not use API data to train its models by default unless your organization opts in.

OpenRouter Decisions offers decision models through your OpenRouter account. The Model picker shows the current choices, and the provider description counts them. All choices support primary or backup checks using your OpenRouter API key. Models and their availability can change, including those marked Experimental. Your OpenRouter account and the selected model's host determine the applicable data controls and terms. Review the provider details before enabling a model.

Removing a model from the picker does not change a connection you already saved. If your selected model becomes unavailable, choose another model, test the connection, and save it.

Choosing None disables Policy Rules and stops rule evaluator requests. Your rules stay saved. Choose a rule evaluator again to resume checks, subject to your plan and connection status.

Backup provider

Backup provider is an optional provider connected with your own key. Any supported customer-key provider can be primary or backup. The backup has its own credentials and connection test. Tap Backup Provider to open the provider list; None means no backup is configured. Choose a provider, configure and test it, then tap Save to return to the list with that provider selected.

Reopen the backup to change its settings or use Test Connection in the Test section below the connection fields. As with the primary provider, unchanged settings test the saved connection; edits are tested before you save them. Select None to remove the backup. Your Policy Rules permission covers the disclosed use of the provider you choose, under your own account terms.

Agent Approve Fast uses Jev first. SubQ Snap handles actions beyond Jev’s input limit under the default oversized-action setting. Perplexity Decider through OpenRouter can check actions when an included provider is unavailable or busy, provided the action fits and enough time remains. All included providers use your AI credits. Your Backup provider is separate: it can check the action when the included service cannot, including when included credits run out. With your own primary provider, the configured backup follows that primary. Your provider bills checks using your key directly; they do not spend AI credits.

If the backup also cannot check the action, If rules cannot be checked applies, except when its own oversized-action setting specifies a different response. You can remove the backup and its key. Withdrawing Policy Rules permission stops all rule checks and generation while retaining saved rules.

If action is too large

Each rule evaluator has an input size limit. Agent Approve Fast defaults to sending an action that is too large for Jev to SubQ Snap. Perplexity Decider through OpenRouter can handle it if SubQ is unavailable and the action fits. Each provider has its own input limit. Your configured backup may handle the action if the included service cannot, within the capacity of its own model. With your own primary provider, the default uses your configured backup, or truncates when you have none. An unavailable backup does not silently switch to truncation; your failure setting applies.

You can change If action is too large:

  • Use backup sends the full action to an included provider that can check it for Agent Approve Fast, then to your configured backup if needed. Other primary providers use your configured backup. Each provider has its own input limit. If no available backup can check the action, the failure setting applies. A primary with no included or configured backup uses truncation.
  • Ask me requests your approval without sending an oversized check.
  • Use Default Behavior applies the active policy's ordinary default behavior.
  • Truncate action sends the beginning that fits. The rule evaluator receives a note stating how many bytes of the original action were sent and omitted. Event details identify the partial check.

Explicit Ask me, Use Default Behavior, and Truncate action choices take precedence over sending an oversized action to a fallback. These choices are independent of If rules cannot be checked. Your backup connection has its own oversized-action setting; it cannot route to another backup.

Truncation applies to any action, including commands, edits, file writes, and tool input. It changes only the rule evaluator input, not the action the agent is requesting. Omitted content is not checked. The rule evaluator considers the supplied prefix and truncation note. Deny and Ask matches can still decide, but a partial check cannot establish that an Allow rule covers the whole action. If no Deny or Ask rule decides, your failure setting applies. Truncation does not guarantee a lower match probability. If even the required context cannot fit, the check fails and your failure setting applies.

If rules cannot be checked

A brief capacity refusal before any evaluation is sent is retried by updated integrations. It does not apply your failure setting or charge credits for an unsent check. If retries are exhausted, the integration reports that it could not complete the request.

Choose what happens after a completed evaluation attempt ends in a timeout, provider error, missing evidence, or exhausted included credits when no backup can complete the check:

  • Ask me requests your approval. It does not auto-approve when the wait expires.
  • Use Default Behavior applies the active policy's ordinary fallback. If that policy uses an auto-approval window, an unchecked action can be allowed when the window ends.

A rule that already matched as Deny still denies. Rules stay saved. To stop a rule from affecting approvals, disable or remove it. No rules are checked when None is selected, your plan does not include Policy Rules, or Allow All or Deny All is active.

AI credits and your own provider

Open Settings > AI Credits to see your allowance, remaining balance, pending usage, and reset date. The usage breakdown shows credits used by Rule Checks and Rule Generator. These are settled charges; pending credits are shown separately. Other appears only when part of the used total has no feature breakdown. Plan & Usage also shows event and active-session limits.

Agent Approve Fast checks and Rule Generator share the included AI-credit allowance. Larger checks, additional included attempts, and generation can use more credits. Rule-check and generator-turn counts show activity; they are not additional monthly allowances when AI credits are enforced.

Pending credits are reserved for work in progress. Unused reserved credits return when that work settles. Paid credits reset at the beginning of each calendar month at 00:00 UTC. Trial credits are a one-time allowance and do not reset during the trial.

When included credits run out, rule checks use your configured customer-key backup if available, then your failure setting. Rule Generator pauses until credits are available. Saved rules and ordinary manual approvals remain available.

When your balance is nearly exhausted, Plan & Usage and AI Credits show Add Backup Provider (or Manage Backup Provider if one is saved). Paid Solo subscribers also see Upgrade to Pro. The comparison shows both plans' allowances and opens Apple's purchase confirmation with Pro selected. Changing plans does not refill a trial's one-time credits.

Checks using your own provider account spend no Agent Approve AI credits. Your provider's billing, usage limits, and data terms apply. An eligible subscription or trial that includes Policy Rules is still required; your own key does not bypass the plan's event or session limits. Rule Generator uses the included service, even when your rule evaluator uses your own key.

What leaves your device

A rule evaluator receives the current action's tool name and arguments when it reaches a rule check, plus the context enabled for that check. It does not read files from your computer or inspect earlier commands. If you choose truncation, it receives only the supplied prefix and the truncation note. Actions decided by your lists are not sent to the rule evaluator.

Agent Approve's provider accounts have zero data retention (ZDR) enabled for Agent Approve Fast rule checks. Fast uses Jev through Agent Approve's TypeSafe account. SubQ Snap runs under Agent Approve’s ZDR agreement. Perplexity Decider runs through OpenRouter with ZDR routing required and provider data collection denied. SubQ does not retain request or response content under that agreement; operational metadata such as model, token counts, and timing may be retained. Prompt and completion logging is disabled on our OpenRouter account. TypeSafe states that it does not train or fine-tune on these inputs. Our OpenRouter account uses ZDR only for the included Perplexity model, excluding Safety Retention endpoints. Its ZDR policy describes how those requests are handled. See our Privacy Policy for the current disclosure. These controls concern provider handling of rule checks. They do not change Agent Approve's event-history retention or temporary evaluation cache. A backup you connect uses your own provider account and its data terms.

Your own provider connection follows your agreement with that provider. OpenRouter also forwards requests to the selected model's host. Review the provider disclosure during setup and the Privacy Policy before enabling a connection.

Rule Generator

Settings > Policy Rules > Rule Generator identifies the providers that draft rules for Generate New Rules. It uses the included AI-credit allowance. It only proposes changes; nothing is added until you tap Add Rules to Policy.

Read a rule check in event details

Open an event to see which rules matched and their match probabilities. Rule Evaluator identifies the provider used for that check. Rule Evaluation Model names its model, separately from the coding agent's model. Rule Evaluation Result shows whether the result was cached, and Provider Response Time and Rule Processing Time show the recorded provider and processing durations in milliseconds. These fields appear when the event includes them.

If a backup checked the action or the input was truncated, the event includes that explanation. AI Credits Used shows the settled charge when available.

What Policy Rules does not do

  • It does not check actions your Deny, Allow, or Ask List already decided.
  • It does not guarantee that any action, including prompt injection or a destructive command, is detected. A rule can miss an action, and a rule evaluator can return a low match probability for an action you would have stopped.
  • It does not read files from your computer or look at earlier commands to check a rule. Each check sees the current action only.

Go out and grab a coffee

We'll ping you if your agents need you. Try Agent Approve free for 7 days.