Your provider keys
By default, runs execute on Coldcache's own provider accounts with Anthropic, OpenAI and Google — that is what a top-up buys, and it is how every workspace starts.
Enterprise workspaces can run on their own provider keys instead. That key is entered in the app and stored in your workspace's own database — on a self-hosted or local install, that is your database on your infrastructure — and requests go directly from your deployment to the provider, with no Coldcache-operated account in the path.
Keys are write-only in the interface: once saved, the UI shows a masked fingerprint, never the value.
What a run stores
The input text and its hash, the version id, the provider and both the requested and resolved model ids, every tool call, every cited source with its URL, title, fetch timestamp and content hash, the output text and its hash, and telemetry — latency, tokens, cost.
That is deliberately a lot. The run record is the evidence, and an audit that arrives eight months later cannot be served by a scrollback.
What is never deleted
Agent versions, runs, inputs and outputs are never hard-deleted. Deprecating a version hides it from new chats and changes nothing about the chats already pinned to it.
When retention expires, content is redacted and the hashes, configuration and telemetry are kept. A diff can still tell you whether something changed after the text itself is gone.
Discarding a draft is the only destructive action in the product, and a draft never had a version number to begin with.
The audit log
Publishes, stable-pointer moves and deprecations are appended with an actor, a timestamp and the detail of the change. The log is append-only and exportable.
It answers the question that mutable assistants cannot: who changed the behaviour of this agent, when, and to what.
What we do not claim
We do not claim the model is deterministic. We claim the causal input set is identical and that everything on our side of the line is frozen.
We do not claim a page you cited still says what it said. We claim we recorded what it said and when, with a content hash, so you can tell the difference.
We do not claim SOC 2 or ISO certification on this build. A report is available under Enterprise terms; asking is the right move if a certificate is a gate for you.
Read the run record before you trust it.
Open the app, run one question, and look at what gets stored. It is all visible in the interface.