> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dexi.net/llms.txt
> Use this file to discover all available pages before exploring further.

# Dexi Guard

> Share your notes with an assistant without handing over card numbers, pay figures or credentials

Dexi Guard gives a connected app a **de-identified view** of your notes. The app reads a copy of each note in which restricted values are replaced by placeholders, and it searches that copy too, so it can't confirm a value it isn't shown. Your own view of your notes in Dexi never changes.

```text What you wrote theme={null}
Offer call with Dana about Falcon
salary:: 185000
Card on file 4111 1111 1111 1111
Staging key AKIAIOSFODNN7EXAMPLE
```

```text What a guarded app reads theme={null}
Offer call with Dana about <DENY_TERM_1>
salary:: <COMPENSATION_1>
Card on file <CREDIT_CARD_1>
Staging key <API_KEY_1>
```

The same value gets the same placeholder throughout a note, so an assistant can still tell two cards apart and follow one through the text.

## Turning it on

Guard is set **per connection**:

* **Settings → Connected apps** — flip the **Guard** switch beside an app you've already connected (the same switch is in **Edit access**). It applies on the app's next request; nothing needs reconnecting.
* **The consent page** — when you connect a new app, the same switch is offered before you approve.

A team owner or admin can also **require Guard for everyone in the team** (Settings → Dexi Guard). Every member's connected apps are then guarded, and members can't switch it off.

Dexi Guard is part of the Team plan. If a plan lapses or changes, connections that are already guarded stay guarded.

## What is hidden

The built-in policy (**Standard**) replaces:

| Kind | Examples |
| - | - |
| Payment and bank details | card numbers, IBANs, routing numbers, labelled account numbers |
| Government identifiers | US Social Security numbers and ITINs |
| Credentials | API keys and tokens that identify themselves by their format, from about seventy services (AWS, GitHub, GitLab, Slack, Stripe, OpenAI, Anthropic, Google, npm, PyPI, SendGrid, Twilio, DigitalOcean, Hugging Face, Shopify, Notion, Linear, Telegram, Dexi and more), Slack, Discord and Teams webhook URLs, access and bearer tokens, private keys, `password = …` assignments in any style (`.env`, YAML, JSON) including longer names such as `SESSION_SECRET`, a password stated in a sentence ("the wifi password is …"), passwords inside connection strings, a random-looking value after "API key:" or "token:", and a long random string standing on its own |
| Codes and recovery secrets | PINs and door, alarm and gate codes beside their label ("PIN 2580", "alarm code is 4471#"), safe combinations, two-factor setup keys, the codes listed under "recovery codes" or "backup codes", and 12 to 24 word wallet seed phrases beside "seed phrase", "recovery phrase" or "mnemonic" |
| Pay figures | a money amount on the same line as a pay word (`salary`, `bonus`, `comp`, `OTE`, `per year`, …), in a table column headed by one, or in the list under a pay heading |
| Sensitive properties | the value of a [property](https://docs.dexi.net/properties-and-bases.md) line whose name says what it is: `salary::`, `bonus::`, `ssn::`, `card::`, `iban::`, `password::`, `api_key::`, … |

**Strict** adds email addresses and phone numbers (international numbers written with a `+`, US and UK national formats, and unbroken digits beside a word such as "mobile").

Free-standing numbers are only treated as a card, IBAN or routing number when their checksum is valid, and a card must also start the way an issuer's cards do, so order numbers, version strings and device IMEIs are left alone. Cards and IBANs are recognised however they are spaced or cased.

> [!WARNING]
>
> Guard finds values by their shape and the words around them. It does **not** detect people's names, postal addresses or company names, and it can't recognise a confidential fact stated in plain prose. A key with no recognisable format is hidden when it sits beside a name such as `token =` or `API key:`, or when it is at least 32 characters mixing upper case, lower case and digits. A shorter unlabelled secret, or a code written with no word saying what it is, is not detected. Use blocked words and property rules for anything specific to you, and treat Guard as a safeguard, not a guarantee.

Pay figures err on the side of hiding: an amount next to a word like "offered" or "per year" is replaced even when it isn't a salary. The exception is a line that reads as a price or a budget. When the only cue is a rate ("per year", "annual", "/hr") and the line names something commercial (a plan, a budget, revenue, a price per seat), the amount is left alone, so "\$96 per year per seat" stays readable. A line that also says "salary", "earns" or "offered" is still hidden.

## The team policy

A team owner or admin shapes the policy in **Settings → Dexi Guard**. It governs every note that belongs to a member of the team.

* **Level** — Standard or Strict.
* **Always hide these words** — client names, project codenames, anything else. Matched as whole words, ignoring case.
* **Property rules** — match a property by name (`*` matches anything, e.g. `client*`) and choose what a guarded app gets:
  * *Hide the value* — a placeholder.
  * *Replace with a code* — a stable code such as `h_3f9a…`. The same value always gives the same code within your team, so an assistant can group or match on it without reading it.
  * *Always show* — exempt a property the built-in rules would hide (say, `salary_band`).
    A value a rule hides is hidden **throughout that note**, not only on its property line. With a rule on `client*`, a note carrying `client:: Halvorsen Logistics` also has "Halvorsen Logistics" and "Halvorsen" replaced in its title and text, by the same placeholder or code. Each capitalised word of a name is covered on its own ("Greta" for `contact:: Greta Halvorsen`); generic words such as "Logistics" or "Group" are not, and neither is a nickname or a name the note never puts in a property. By default this reaches one note at a time: another note that mentions Halvorsen without the property is left as written.
* **Also hide a rule's values** — *Elsewhere in the same note* (the default), or *In every note in the library*. With the library setting, a name hidden by one of **your own** rules in any note is hidden in every note the policy governs, yours and your teammates', by the same placeholder (or, for *Replace with a code*, the same code, so an assistant can follow one client across notes without reading the name). Notes that mention a newly added name are re-checked first and are not visible to guarded apps until they have been. Only names are shared this way: numbers, and the built-in pay, card and password properties, stay within their own note. A first name that is also an ordinary word ("Will", "Grace") is hidden wherever it is capitalised. When the property is removed, notes checked while it existed keep the name hidden until they are next edited.
* **When a note has restricted values** — replace them with placeholders (the default), or hide the whole note from guarded apps.

The **preview box** on the same page shows any text the way a guarded app would receive it, using the policy as you have it on screen, before you save.

If you belong to someone else's team, that team's policy governs your notes; otherwise your own does. Without a saved policy, the built-in Standard policy applies.

## How a guarded connection behaves

* **Reading** — every tool that returns note content returns the de-identified copy. Items carry `"guarded": true`, and `redactions` when anything was replaced (see [the note item](https://docs.dexi.net/mcp/tools.md#dexi-guard)).
* **Searching** — keyword, regular-expression and semantic search, similar-note lookup, tag filters and tag counts all run on the de-identified copy. A search for a hidden value finds nothing.
* **Writing** — creating a note and appending to one work normally: what the assistant writes is stored as written, and the response shows it de-identified. **Replacing** a title or body that contains hidden values is refused, because writing back the placeholders would overwrite the originals; append instead.
* **Bases** — `query_base`, `get_base`, `create_base`, `update_base` and `delete_base` are not available, since a base filters and summarises on the real property values. `list_bases` still answers.
* **Not yet checked means not visible** — a note appears to guarded apps only once its de-identified copy exists. Notes you write or edit are checked within moments, and ones that arrive by feed, email or the web clipper within a few minutes; turning Guard on for the first time, or changing the policy, re-checks the whole library in the background, and each note reappears as it's done. Settings → Dexi Guard shows the progress.

## What Guard does not cover

Guard applies to apps connected over MCP. It does not change:

* what **you** or your teammates see in the Dexi app,
* share links and [published feeds](https://docs.dexi.net/api/feeds.md), which show a note as written to whoever you give them to,
* exports and digest emails, which go only to the note's owner.

Folder names, base names and people's display names are passed through a final check on the way out, but are not otherwise rewritten.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.