> ## Documentation Index
> Fetch the complete documentation index at: https://docs.splox.io/llms.txt
> Use this file to discover all available pages before exploring further.
# Data and privacy
> What Splox stores, where it lives, how long it stays, and who can see it
Three places hold your work: the database, the machine's disk, and the git
repository behind the harness. This page says what goes into each, what is kept
and for how long, and who besides you can read it.
The legal document is the [privacy policy](https://splox.io/privacy); Splox LTD
is registered in Bulgaria and the policy names the sub-processors and your
rights under GDPR. What follows is the operational half — what the running
system actually does.
## What is stored, and where
| | |
| ----------------------- | ---------------------------------------------------------------------------------------------------------------- |
| **Chats** | name, which machine and project, whether it is shared, timestamps |
| **Messages** | every turn, both sides, with the attachments you sent |
| **Tool calls** | the call, its arguments and its result, as the transcript shows them |
| **Runs** | status, timings, the harness commit that ran, the model and endpoint, failures |
| **Run events** | the journal a stream replays from |
| **Usage** | one row per model call: provider, model, token counts, latency, cost. No message text |
| **Billing** | transactions, balances, subscription state, the windows |
| **Account** | email, name, sign-in identities, sessions |
| **Connections** | MCP servers and provider endpoints, with their credentials encrypted (AES-256-GCM) and never returned by any API |
| **Environment secrets** | encrypted the same way, decrypted only to be injected into your sandbox |
| **API keys** | stored as a SHA-256 hash; the plaintext is shown once, at creation, and is not recoverable |
| **Consent** | which cookie choice was made, when, and from which browser |
Everything the agent writes: files, checkouts, databases, logs, installed
packages, whatever a program left behind. It is a disk on the runner, it is not
indexed or read by the platform, and it survives a stop.
A machine that has been stopped for 30 days is archived: the filesystem moves to
object storage and comes back on the next start. Nothing is lost by it.
Deleting the machine destroys the sandbox and the disk with it.
The harness is a git repository — programs, agents, hooks, tools, evals, the
docs it carries — with every published version as a ref and its full history.
Anything the agent commits into `~/harness` and publishes is stored there
permanently.
A workspace pushed with `publish_workspace(path)` lands in a repository of its
own, on the same host.
## How long it stays
Short version: chats, runs and their events are kept until you delete them.
There are exactly two scheduled sweeps on the platform, and neither touches
anything you wrote:
| Runs | What it removes |
| ----------------- | ---------------------------------------------------------------------------------------------------------------------------- |
| hourly | notification rows older than six hours — content-free pokes that tell a connected client "something changed", never messages |
| every ten minutes | an internal pass over tool errors, for improving the tools |
No job deletes a chat, a message, a run or a usage row on a schedule. Nothing
expires on its own. If you want a conversation gone, delete it; if you want all
of it gone, delete the account.
The API's event journal can answer `410 event_cursor_expired` for a stream cursor
that points before the retained window. It is part of the streaming contract —
resume without a cursor and you get the whole run — and not a statement that the
run's events have been removed.
## Who can see it
**You**, through the app and the API, with your session or an API key.
**Anyone holding a share link.** A shared chat is a public URL with no sign-in:
the token is the access control. Making the chat private kills the link
immediately, and sharing it again mints a different one. [Security](/security)
has the detail.
**The model provider behind the run.** The text of a turn, its attachments and
the tool results in its context go to whichever endpoint the run resolved to —
Splox's account with that provider on a platform endpoint, or your own account
on a connection you added. Before the first time this happens, the app asks:
> **Allow Splox to send your inputs to AI providers**
>
> To generate responses and run agents, Splox sends your messages, attachments,
> files, prompts, agent context, and related inputs to third-party AI providers
> such as OpenAI, Anthropic, and Google, and to tool integrations you enable.
Declining means the message is not sent and the action does not proceed.
**Servers you connect.** A tool call sends what that call needs to the MCP
server it belongs to, with the credential you connected it with.
**Nobody else.** A chat, a machine and a harness are yours; another account's
request for one of them answers `404` rather than `403`, because telling a
prober that something exists is itself an answer.
The privacy policy is where the sub-processors are named and where the transfer
basis for processing outside the EEA is set out. Read it if any of that decides
whether you can put a piece of work here. It makes no statement about your data
being used to train models — do not assume one in either direction; ask
[support](/support) if it matters to your case.
## Analytics and cookies
Essential cookies keep you signed in and protect the session; they cannot be
turned off. Product analytics are a separate, opt-in category — off unless you
say yes — and the choice is stored on the device and logged on the server as the
record that it was asked. **Reset consent** brings the banner back.
## Getting your data out
There is no export button. What there is:
* **A chat** — the transcript is on `GET /api/v1/chat-history/{chat_id}/paginated`.
* **Runs** — [the v2 API](/api/runs) returns a run's messages, outputs, events,
usage and its tree of sub-agent runs.
* **The harness** — `git clone` the repository; it is an ordinary git remote and
the history comes with it. See [Publishing a version](/inside/versions).
* **Files on the machine** — the agent can hand you any of them as a download
link, or push a whole directory to a git remote you can clone.
## Deleting
Deleting a **chat** removes it from the product: it stops being listed, it 404s
for you, and its share link 404s for everyone. The row is marked deleted rather
than erased.
Deleting a **machine** destroys the sandbox and its disk. Deleting a **harness**
destroys the repository and every machine running it.
Deleting the **account** is the erasure. **Settings → Privacy → Delete account**,
confirmed by typing the phrase:
> This permanently deletes your account and all associated data — harnesses,
> agents, chats, credentials and billing history. Any active subscription is
> cancelled immediately.
Every sandbox you own is destroyed, Stripe is told to cancel and to delete the
customer, and one database transaction removes the rows: chats, messages,
tool-call records, runs and their events, machines, MCP servers, skills, push
tokens, webhooks, sessions, identities, balances, transactions, subscriptions and
usage. It is not reversible and there is no grace period.
Two things are kept on purpose: the consent log, with your user id set to null,
as the evidence that consent was asked for; and the git repositories, whose
history has never been removed by deleting a user's rows.
To exercise a GDPR right that this page does not cover — access, rectification,
portability, restriction, objection — the policy names `support@splox.io` and
undertakes to answer within 30 days.
Where the work happens and what can reach it.
The legal document, with the sub-processors named.