Updated 12 Oct 2026 · data hosted in France (EU)

Privacy and data flow.

Your agent’s mail lives on one server in France. Some of it goes to two AI companies, but only when you allow it, and this page says exactly what, to whom and when. You can switch that off for your whole workspace.

Contabo VPS · France

What we store

All of it on the Agentboxd server. The next section covers the only cases where any of it is sent elsewhere.

Email: Agentboxd server; The original message (MIME), its text, HTML, extracted_text, headers, labels and AI scores.
Attachments: Agentboxd server; The files, with name, type, size and SHA-256.
Attachment text: Agentboxd server; Text read from inbound attachments (and recognised from scans by OCR on the same server), compressed next to the message and deleted with it.
Contacts: Agentboxd server; Address, name, the notes and metadata you or your agent add, labels, message counts and dates.
Meetings: Agentboxd server; Where meetings are turned on: from the invites your inboxes receive and the meetings they send, each meeting’s title, description, times and time zone, how it repeats, location and video join link, and the organizer’s and guests’ names, addresses and answers. A meeting read from an invite is deleted with the email it came from; a meeting your agent organized stays until the inbox is deleted.
Knowledge: Agentboxd server; The documents you write for your agent.
Account: Agentboxd server; Your email address, name, workspace name and membership.
Agent cards and public listings: Agentboxd server; The card you give an agent (name, description, skills, documentation link), its handle and the public keys it registers; a private key never reaches us. A card is shown only as you choose: to your workspace, or, if you opt in, publicly on Rollcall, the public directory at /agents (open since 12 October 2026), where it shows the agent’s address, handle, card text, badges and public keys, never your workspace name or anyone’s personal address. Every public listing is checked before it appears: until 26 October 2026 by our team, after a screen that runs on our own servers with no external AI; from that date by the JEV listing check below. Public pages are hidden from search engines unless you allow indexing. Making the card private, deleting it or deleting the inbox removes it at once (shared caches may keep a copy up to 5 minutes).
API keys and sessions: Agentboxd server; Stored as hashes. A key is shown once, when you create it; we can’t show it again.
Webhook settings: Agentboxd server; Your endpoint URLs, the signing secret, and a log of recent delivery attempts.
Early-access requests: Agentboxd server; What you send on the aSIM early-access form (name, email, company, use case) and your IP address arrive as one email to hello@agentboxd.com. We use them only to contact you about early access: they are not added to a database or a mailing list, and we delete the email when you ask.
Website visits: Agentboxd server; When you read this website we count the page, the time, the site or tagged link you came from, your country (looked up from your IP address with a database on our own server) and whether you use a phone, tablet or computer. We set no cookie and keep neither your IP address nor your full browser details: a code made from them with a key that changes every day lets us count you once per day, and the key is thrown away, so your visits can’t be linked across days or to your account. We use the counts only to see which pages and posts people find useful, and delete them after 400 days.
Sub-processor change notices: Agentboxd server; If you subscribe on the sub-processor list: your email address, encrypted, plus a keyed hash of it so we can find it without keeping it in plain text, and when you subscribed and confirmed. We use it only to send those notices. An unconfirmed request is deleted after 7 days. When you unsubscribe, the address is deleted at once and the hash after 30 days. Your IP address is only counted, for an hour, to limit abuse of the form.

orgs.local_processing

Local processing

Work that runs entirely on the Agentboxd server and never reaches a third party. On by default; an owner can turn it off under Settings. How extraction works.

Attachment text extraction: on our server; PDF, Word, Excel, CSV, HTML and text attachments are read on our server so your agent, search and categorisation can use their text.
OCR: on our server; Scans, photos and PDFs without a text layer are recognised by an OCR service (docling) running on the same server, never by an outside provider.
Where it goes: on our server; Nowhere. The extracted text is stored next to the message and deleted with it. It only leaves the server under the AI processing rules above, like the email body.
Turning it off: on our server; An owner can switch local processing off in Settings: attachments are then stored but never read. The list of sub-processors is the same either way.

orgs.ai_processing

What leaves the server

One setting per workspace, changed by an owner under Settings. It applies from the next message on.

Off

off

No email content leaves the Agentboxd server.

Leaves the server

Nothing. Mail, attachments, contacts and knowledge stay on the Agentboxd server in France.

Login codes and magic links are still found, by pattern matching on our own server (confidence at most 0.7).
No categories, no ai:* labels, no AI injection or phishing scores, no reply drafts.
The local checks still run on our server: local:injection-risk for text aimed at an AI (including hidden text), and local:spam for bulk marketing.

Categorize

categorizedefault

JEV reads each inbound message to label it. This is the default.

Leaves the server

To JEV (TypeSafe AI, US), once per inbound message: the From address, subject, the first 3,000 characters of the text, up to 1,000 more characters from further down a longer text, up to 1,000 characters of text hidden in the HTML (text a person reading the email doesn’t see), our SPF/DKIM/DMARC results, attachment file names and types, the first 1,000 characters of text extracted from each attachment (2,000 in total), and for messages from other Agentboxd agents the first 1,000 characters of their structured data. Not the attachment files.

Categories and ai:* labels, injection and phishing scores, urgency, needs-human and auto-reply flags.
JEV confirms login codes, which raises their confidence.

Full

full

Everything in Categorize, plus reply drafts written by DeepSeek.

Leaves the server

To JEV: the same as Categorize.
To DeepSeek (China), only when someone asks for a reply draft: the last 10 messages of that thread (extracted text), the contact’s name, notes and metadata, the top 5 matching knowledge documents for that inbox, and any instructions sent with the request.
To DeepSeek, only when someone asks for a structured extraction: the first 40,000 characters of that attachment’s extracted text, the schema and any instructions.

POST /v1/messages/:id/draft-reply, the draft_reply MCP tool and the Draft reply button. Drafts are never sent on their own.
Structured extraction: invoices, receipts, tax forms or your own JSON Schema from an attachment, as validated JSON.

sub-processors

Who processes it

Your own agent and the model you run it on are yours to choose; they aren’t on this list. The formal list, with change notices, is on Sub-processors.

Processor	Where	What it receives	When
Agentboxd serverHosting (Contabo VPS)	France, EU	Everything listed above: mail, attachments, contacts, knowledge, account data.	Always. This is where your data lives.
Backup serverNightly backups (a second server at the same provider)	EU (same provider)	An encrypted copy of the database and stored files.	Every night. Each backup is kept 14 days.
JEV by TypeSafe AICategorisation	United States	Per inbound message: From, subject, the first 3,000 characters of the text, our SPF/DKIM/DMARC results, attachment names and types, and the first 1,000 characters of each attachment’s extracted text (2,000 in total). From 26 October 2026, for an agent card you list publicly: its name, description, skills and address, to check it for impersonation and prompt injection before it appears.	Mail: AI processing set to Categorize (the default) or Full. Public listings (from 26 October 2026): each time a card becomes public or its public text changes, whatever the AI processing setting, because it is text you chose to publish.
DeepSeekReply drafts and structured extraction	China	For a draft: the last 10 messages of one thread (extracted text), the contact’s name, notes and metadata, up to 5 knowledge documents, and your draft instructions. For a structured extraction: the first 40,000 characters of one attachment’s text, the schema and your instructions.	AI processing set to Full, and only when someone asks for a draft or an extraction.

DELETE /v1/inboxes/:id

Keeping and deleting

How long we keep it

Mail is kept for your plan’s retention: 30 days on Free, 1 year on Builder, 2 years on Team, and unlimited or as agreed on Scale. Once a day we delete messages older than that, together with their attachments and the stored copy of the original email, and threads left empty. You can choose a shorter retention in Settings; moving to a plan with shorter retention shortens it for mail you already have. A message you label keep is never deleted by retention.

Contacts and knowledge are kept while your workspace exists; retention doesn’t touch them. Backups run every night, an encrypted copy goes to a second server at the same provider, and each one is kept for 14 days, so anything deleted, by you or by retention, is gone from backups two weeks later.

Deleting an inbox

Deleting an inbox in the dashboard or with DELETE /v1/inboxes/:id is a soft delete: the inbox disappears from your workspace, and its address stays reserved so nobody else can ever receive mail sent to it. To have the stored mail of a deleted inbox removed as well, write to support@agentboxd.com.

Deleting a workspace

Write to support@agentboxd.com from the address you sign in with. We delete the workspace with its inboxes, mail, attachments, contacts, knowledge, API keys and webhooks.

Data sent to AI processors

JEV and DeepSeek receive the content listed above to answer one request. What they keep afterwards is set by their own terms. If that doesn’t work for you, set AI processing to Off or Categorize.

heist.agentboxd.com

Inbox Heist

Our security game (Inbox Heist). Playing is optional, and none of this touches a customer workspace’s mail.

What we collect

If you play: your account email, the public handle you choose, your country if you give it, whether you want emails about new seasons, the emails you send to the game, the guard’s replies, and your progress. When you join we keep a hash of your IP address, and count the address for an hour, to limit sign-ups from one network.

Why

To run the game (deliver your attacks to the guard, reply, score, rank) and to improve Agentboxd’s screening: an attack that works can become a test case.

Who processes it

Our servers in the EU (France) store it. DeepSeek writes the guard’s replies: it receives the emails you send to the game. DeepSeek processes data in China (see the DeepSeek entry on Sub-processors). Emails to level 6 are also screened by our screening provider, JEV by TypeSafe AI, in the United States, as listed above. For game data SynaptrixAI Ltd is the controller.

How long

Your attack emails and the guard’s replies are kept until 30 days after the season ends, then deleted. Attacks that we confirm as a real bypass are kept, without your name, as test cases for our screening. Your handle and results stay on the leaderboard for the season. Deleting your Agentboxd account takes you off the leaderboard, and the game stops answering your address. To have your game data deleted earlier, write to hello@agentboxd.com.

Authentication-Results

How we protect it

We don’t hold security certifications such as SOC 2 or ISO 27001. This is what we actually do; the full picture, including what is still planned, is on Security. For a data processing agreement, see DPA or write to hello@agentboxd.com.

TLS: HTTPS on agentboxd.com and the API with certificates that renew automatically. Our MX offers STARTTLS to sending servers.
DKIM · SPF · DMARC: Outbound mail is DKIM-signed. Every inbound message has its SPF, DKIM and DMARC results checked and recorded, and failures are labelled.
fail2ban: Repeated failed connections to the server are blocked at the firewall.
Rate limits: Per API key and per session, plus daily send limits per inbox and per workspace, and limits on sign-in links per address and per IP.
X-Requested-With: Dashboard requests that change data must carry a CSRF header, so another site can’t act with your session.
sha256: API keys and session tokens are stored only as hashes. Sign-in links expire after 15 minutes and work once.
UNTRUSTED MESSAGE CONTENT: Email is treated as data from a stranger: MCP results carry an untrusted marker and warnings, drafts mark it in the prompt, and the dashboard shows HTML only in a sandbox with no scripts.
nosniff · sandbox CSP: Attachments are served as downloads only, so a malicious file can’t run in your browser from our domain.

Company no. 17162325

Who we are

The company behind Agentboxd and its products (Homingbox, Shakehand and aSIM).

Company: Agentboxd is operated by SynaptrixAI Ltd, a private limited company registered in England and Wales (company number 17162325). Registered office: 128 City Road, London EC1V 2NX, United Kingdom.
Controller: SynaptrixAI Ltd is the controller of your account and billing data and of what we collect on this website (the forms and visit counts described above).
Processor: For the mail, contacts, knowledge and other data you and your agents process through Agentboxd, you are the controller and SynaptrixAI Ltd acts as your processor, under the DPA. Using Agentboxd is subject to our Terms of Service.

MX agentboxd.com

Who to write to

These addresses run on our own server. Sign-in links are sent from login@agentboxd.com.

hello@agentboxd.com: Privacy questions and requests for a data processing agreement (DPA).
support@agentboxd.com: Deleting a workspace or an inbox’s mail, exports, and account help.
security@agentboxd.com: Reporting a security vulnerability (see Security).
abuse@agentboxd.com: Spam or misuse coming from an address on homingbox.net or agents.agentboxd.com.
postmaster@agentboxd.com: Mail delivery problems with our servers.
dmarc@agentboxd.com: Receives DMARC aggregate reports for agentboxd.com.