Privacy and the install report

The short version

• The running platform never contacts Abada. The Engine, Studio, the Agent Worker and Keycloak talk only to what you configure or enable: your database, your identity provider, your model providers (to answer agent steps and, once a day, to list the models your key can use), your MCP tool servers, the public model price catalog (once a day, on by default) and, if you turn it on, your own telemetry collector. Telemetry stays off by default and goes only to your collector. None of these calls carries data from your processes unless it is a model or tool call your process makes.
• The Linux and macOS installer sends one anonymous report when it finishes, so we can tell how many installs succeed or fail and where they fail. You can switch it off.
• The download server keeps a short log of downloads: which file, when, which release and the coarse location of the request. It cannot be switched off for a download, because it is part of serving the file.

Nothing here includes your processes, your data, your keys, your host name, your user name or your IP address.

What the installer reports

install.sh sends a single request to install.abadaplatform.com when it exits, successful or not. The installer prints what it will do before it downloads anything. The report holds only these fields:

Field	Example	Why
outcome	install_succeeded or install_failed	Is the install working for people?
failed step	download, verify, start…	Where installs break
Abada version	1.1.0	Which releases are in use
install id	a random id the installer made; it is kept in .abada-install-id next to the install	Tells one install’s report from another’s
operating system, CPU	Darwin, arm64	Which platforms to test
Docker version	27.3.1	Which Docker versions to support
profile, duration	dev, 212 seconds	How long an install takes
key supplied	true or false	Whether a model key was provided at install (never the key)
CI	true or false	Whether the CI variable was set, to separate automated installs

The location of the install (country and region only) is not sent by the installer. The receiving server derives it from the request’s address when it arrives. The address itself is not stored, and nothing finer than the region is kept.

The server accepts only the fields above and refuses anything else, so the installer cannot be made to send free text, file paths or names. You can read both ends: install/install.sh and install/worker/worker.js.

What the download server records

When the installer script, the version pointer (/latest) or a release archive is requested, the server records: the time, what was requested (the script, the version check or the archive and its release) and the coarse location (country and region) derived from the request. Nothing else is taken from the request: no address, no hash of it, no user agent.

What is never recorded

Your IP address (or any hash of it), city, network owner, host name, user name, file paths, environment variables, API keys, process definitions, database contents and anything you do after the install. The platform’s own Docker images are pulled from the GitHub Container Registry, which, like any registry, sees those pulls; Abada does not.

How long and who sees it

Records are deleted automatically after 13 months. They are used to count installs and adopted versions, to find where installs fail and to decide which platforms to support. They are stored in Cloudflare, which hosts the download server, and are read by the maintainers of the project. Because a record holds no name, no address and no city, it cannot be tied to a person afterwards, so a request to delete “my” record cannot be matched to one.

Connections the running platform makes

Connection	When	What it carries	Turn it off
Your model providers	Each agent step, Insight and authoring call; a test call when a provider is saved; their model list once a day	The prompt and inputs your process declares; for the test, a fixed “Reply with READY”; for the list, only the key	Remove the provider; ABADA_MODEL_DISCOVERY_ENABLED=false for the daily list
Your MCP tool servers	Each tool call an agent makes under its policy	The tool arguments	Remove the tool binding
Public price catalog (openrouter.ai by default)	Once a day	Nothing: a plain GET of the public model list, no key, no identifier	ABADA_PRICE_CATALOG_ENABLED=false (the release snapshot then prices what it knows), or point ABADA_PRICE_CATALOG_URL at a mirror
Your identity provider, database, telemetry collector	As configured	What each needs	Their own settings

Opt out

# For one install, for the whole pipeline (the variable goes to bash):
curl -fsSL https://install.abadaplatform.com/install.sh | ABADA_USAGE_PING=0 bash

# For every tool that honours it:
export DO_NOT_TRACK=1

With either set the installer prints no notice and sends nothing. The installer also sends nothing when ABADA_RELEASE_BASE_URL points to a mirror. To avoid the download log as well, download the release archive from the GitHub release and run ./release/abada-platform up dev from it, or serve it from your own mirror.

Note

Only install.sh sends a report. The PowerShell quickstart, the release archive and a git clone send nothing of their own; the downloads they make from install.abadaplatform.com appear in the download log described above.