Privacy

devdash is an internal tool. It is not open to public sign-up: an administrator decides which accounts may use it.

This page says what this deployment reads, what it keeps, who can see it, and what leaves for a third party. Where something is held that you cannot yet see or remove, it says that too.

Signing in

Signing in with Google requests only openid, email and profile. Signing in alone gives this deployment no access to your mail, your files or your calendar.

It records:

Where an administrator has granted access to an address, the address, who granted it, and when, are also recorded — including after the grant is revoked, so there is a record of who had access and when it ended.

Connecting a mailbox is a separate decision

You can additionally connect your Gmail account, which asks Google for gmail.readonly. This is a second consent screen and a second decision. It is deliberately not part of signing in, because signing in to a dashboard should not hand over a mailbox.

One connection, two different things read your mail, scoped two different ways. Project collection is below. The inquiry sweep is the section after it, and it is the wider of the two.

Project collection

Project collection starts when somebody declares a counterparty on a project you are a member of — an address such as dana@acme.com, or a whole domain such as acme.com. The search then runs at Google: the query names those counterparties and a start date, and the query builder refuses to produce one that names nobody, so mail outside that scope is never sent to this deployment.

For each message that matches, it stores the sender and recipients, the subject, the date, the message body, and which declaration matched it. That last part is there so you can challenge it: every message collected this way can say why it is there.

A sync runs when somebody presses the button on a project page, or posts to the sync endpoint. It reads your mailbox using your connection and nobody else’s. There is no background schedule in this deployment that reads mail on its own.

The inquiry sweep reads mail no project has declared anything about

An inquiry is the first message from somebody no project has declared, so a party-scoped search cannot find it. The sweep is scoped by something else instead, and that makes it the widest read here.

You run it yourself, and you supply the scope each time: one or more shared intake addresses that mail was sent to, or Gmail labels you have applied, plus a start date. The date is never defaulted, the window is bounded in length and in how far back it can reach, and a sweep that names no address and no label is refused rather than run — there is no unbounded sweep and no flag that produces one. Your own address cannot be used as a selector, because to:you@… would select nearly everything you receive. Nothing about the sweep is stored as a setting, so there is no sweep that keeps running after you stop asking for one.

What it reads is everything matching those selectors in that window — not only mail from clients, and not only mail about work. Each message it reads is sent to Anthropic to be judged (see below), and every message it reads is stored, including the ones judged not to be inquiries: the sender, the date, the subject, and the full message body, along with the model’s verdict and its reasoning. A newsletter that arrived at a shared intake address is read, sent, and kept, with “not an inquiry” recorded against it.

What is stored by a sweep is readable by you and by nobody else in this application. You can dismiss a candidate, which is durable — a later sweep over the same message changes nothing, so a dismissal stays dismissed. What dismissing does not do is delete the stored body. As this deployment stands there is no control that removes a swept message, and no period after which its stored body goes: the exclusion guarantee described further down covers your communication record and does not yet reach anything a sweep stored. That is a gap in what this page can promise you, and it is stated as one.

Slack, if it is connected

Slack is read with a user token, not a bot token, and the scopes are search:read, users:read and users:read.email. This is a bigger ask than gmail.readonly and it is worth being exact about why. search:read on a user token searches everything you can see in that workspace: public channels, private channels you are in, group messages, and your direct messages. Slack’s own consent screen does not let you narrow it, and Slack’s search has no parameter that excludes direct messages.

The narrowing to declared counterparties is done by this application’s code, not by the permission you grant. Every query it builds names a declared party — the address is resolved to a Slack user through users.lookupByEmail — and it never lists conversations or walks channel history. So a message with nothing to do with a declared party is not transferred. But the token itself could reach your direct messages, and the only thing standing between it and them is our query builder.

A collected Slack message is stored with the declared party’s address as its only participant. The conversation roster is never stored. Its “subject” is a line this application writes — for example, “Slack message from @dana in #acme-rollout” — rather than anything a person typed, and the author’s handle or Slack user id is kept on the attribution line at the top of the stored text.

There is no screen in this deployment that connects Slack. The collection code runs whenever a Slack connection exists for your account; nothing here creates one yet.

Google Drive, if it is connected

Drive is a separate Google consent from Gmail, and it asks for drive.metadata.readonly. Said plainly: that is read access to the metadata of every file in your Drive, including personal ones — names, owners, who a file is shared with, timestamps, and links. A filename discloses plenty on its own, and this scope covers all of them, not only the ones connected to a project.

The other half, and it is equally load-bearing: no file contents are ever fetched. The narrow scope is chosen precisely so that they cannot be — drive.readonly, which Google describes as seeing and downloading all your files, is not requested, and no field that carries or links to content is asked for.

Searches name one declared address at a time, so the files listed are ones shared with or by a declared counterparty. For each one kept, it stores the file name, its type, whether it is a folder, its owners and the addresses on its sharing list, the created and modified times, when it was shared with you, its webViewLink, and which declared address justified keeping it.

Two honest gaps. There is no screen in this deployment that connects Drive, and no screen that lists what Drive collected — so a Drive record is held where you cannot see it, and the exclusion control described below has nothing to act on for these. They are also absent from the export of your record.

Circleback

A Circleback API key is yours alone, so only meetings you attended are read. Circleback has no server-side participant filter, so unlike Gmail, Slack and Drive the narrowing happens here rather than at the source: meeting records are paged into this deployment and compared against the declared counterparties locally. Only meetings involving a declared counterparty are kept.

Discord

Discord collects nothing. There is no way to connect a Discord account, nothing designates a channel for a project, and the adapter skips every message it sees as unjustified. The panel on the connections page says “not built” for that reason.

If it is built, one limit will still hold: a Discord bot cannot read your direct messages with another person. That is a limit of Discord’s own interface, not a setting.

Nobody can put your mailbox in scope but you

A project’s owner can invite you to a project. They cannot add you to one. Nothing of yours is collected until you accept: until then the project reads no mailbox of yours, whatever is declared on it.

When you are invited, a notification is raised for you. It appears on Notifications and in the count on your home page, and it says what accepting would cause before you answer. Your invitations page lists the counterparties declared on that project now, so you can see what your mail would be searched for. Declining is one click and no reason is asked for.

Nothing is emailed to you. This deployment sends no email at all, so a notification here is the only thing that tells you, and it reaches you only once you sign in. If the address invited has no account here yet, the invitation waits on that page for your first sign-in and no notification is raised for it — that is the one case where nothing reaches you until you come looking.

Once you are on a project, any member of it can declare a new counterparty, and from the next sync your mail with that counterparty is collected. That is the scope you accepted. Every project you are on names its live declarations, who declared each one, and when.

Withdrawing a declaration is prospective. It stops the next message; it does not remove what was already collected, and the interface says so before the click.

What is sent to Anthropic

This is the section worth reading twice. Where this deployment is configured with an Anthropic API key, four different things send text to Anthropic’s API. Where no key is configured, nothing is sent: work-item suggestions are off, and the inquiry sweep refuses to run rather than reporting that it found nothing.

That is a third party receiving the contents of your correspondence and the addresses of your colleagues. Nothing is sold, and nothing is shared beyond this.

Who can see what was collected

Your record is yours. Messages collected from your account are readable by you and by nobody else in this application by default. Every read of the communication record is scoped to one person’s id, and there is no function anywhere that lists a project’s records across employees.

There is one screen that shows you correspondence that is not yours, and it is /search. It has a second, separately labelled section for records a colleague contributed to a communications request you issued. Nothing reaches it except through that route: being an administrator does not qualify, and being on the project does not qualify.

Someone may formally ask for communications about a client or a period, stating why they need it and what they tried first. You then choose whether to contribute, and declining is a recorded answer rather than a failure to respond. Because this is a small team, whether each person contributed or declined is visible to whoever asked — you are told that before you choose, because a count would not anonymise anything at this size and implying otherwise would be worse than naming it.

A preservation hold is different and is labelled as such: participation is required, and the person who issued it cannot quietly withdraw it.

The audit log

Declaring and withdrawing a counterparty, adding and removing a project member, connecting and disconnecting a source, issuing, answering and withdrawing a request, disclosing records, excluding a message and revoking an exclusion, granting and revoking access — each of those writes an entry to an append-only log. The database refuses updates and deletes on that table through a trigger, so a wrong entry is corrected by appending another and the original stays.

The log cannot hold the contents of your mail: its detail field accepts only a fixed vocabulary of names, and subject, body, snippet and participant fields are excluded from that vocabulary by a check the build and the database both enforce.

Two limits. The log is readable by administrators, not by you. And running an inquiry sweep writes no entry to it — that is a gap, not a design.

Removing things

You can exclude any message from your record, and the exclusion is permanent against future collection — the next sync consults the exclusion list before it inserts anything, so it will not put the message back. Excluding removes it here and deletes nothing from your mailbox.

That guarantee covers the communication record. It does not cover inquiry candidates, which can be dismissed but not removed, and it has no reachable effect on Drive records, which no screen lists. Both are described above.

Disconnecting a source stops future collection from it and deletes the access and refresh tokens this deployment holds for it. It does not itself revoke the grant at the provider; for Google, do that at myaccount.google.com/permissions.

For your account itself, ask an administrator. Revoking access takes effect on your next request.

Getting your own copy

You can download your whole record, as CSV or as JSON, from /record. It covers every project you can see: the messages collected for you, the exclusions you have made, and each message’s justification resolved into a sentence rather than left as an id — including justifications under declarations that have since been withdrawn, because that is the record somebody is likeliest to contest.

What the file does not contain: Drive records, and inquiry candidates. Neither is exported today.

Back to sign-in