AI value and governance, without keeping your prompts.
Valistry works with usage, billing and settings metadata. Prompts, replies, transcripts, documents and source code stay outside the ledger. Read-only to start; an agent acts at your vendors only with a key you give it.
- model: claude-sonnet
- promptrefused
- tokens: 1,204 / 388
- completionrefused
- cost: $0.0132
- source coderefused
What lands in the ledger
Six fields
Provider and model, when, how much, what it cost, how it went, whose. Content has no field to land in.
Every field, and what it holds →Only the metadata needed for AI value and governance.
This is the whole shape. A reviewer can check it against the connector code, which we share for review under NDA.
- Provider and model
- Which vendor served the call, and which model
- When
- Timestamp and billing period
- How much
- Tokens, requests, seats or seconds, by the vendor's own meter
- What it cost
- Rate, currency and the reconciled amount
- How it went
- Status and latency
- Whose
- Account, project and the identifiers the vendor attaches
Never stored
The record has no place for content.
These six things cannot enter the ledger, because no field exists to hold them.
- Prompts
- Completions
- Conversations
- Transcripts
- Documents
- Source code
Seven promises, each with its proof.
Pick one to see how the code keeps it.
Access
Read-only first
Each source asks for the narrowest access it supports, and we only read at first. An agent changes something at your vendor only with a separate key or app you give it for its one job, after an owner or admin approves or within the limits you set.
The evidence
Each connector declares the URLs and calls it may make. Our code refuses any other. Every agent change records what it replaced, and a change made for a saving is checked on the bill after. Most can be undone in one step.
The security review, answered.
What the software does today. Connector code is open to review under NDA.
Access
- Sign-in
- Magic link, one-time code or Google, no password. SAML and SCIM tested with Microsoft Entra ID.
- Roles
- Owner, admin and member. Owners and admins make every change. A member reads, asks and keeps their own saved views.
- Least privilege
- Connections start read-only. An agent gets its own key for one job.
Your data
- Never kept
- Prompts, replies, transcripts, documents and source code.
- Encryption
- TLS in transit. Connector credentials: AES-256-GCM under one server key.
- Tenant isolation
- Our application code limits every query to your organization. The database does not enforce this yet.
Where it lives
- Region
- Your records stay in the region you choose. Sign-in lookups, error reports, screen counts and mail go through the US.
- Deletion
- Delete your organization and its records go with it. One row stays: who deleted it, its id and when.
- Sub-processors
- 9, each named on the privacy notice.
Evidence
- Audit evidence
- Tools, owners, seats and readiness, as files.
- Call log
- Vendor calls are logged, unless a write fails. Only deleting your organization erases them.
Bring the metadata. Keep control of the content.
We answer security questions in writing.