⚖

Enterprise Token Governance Platform

Dependable Token governance for the enterprise, and a supply of AI tokens the systems you put into production can rely on.

Talk through your setupRead the docsWhat it does

How Tare works

Tare sits between your application and the model vendors, consolidating vendor selection, protocol differences, failure handling and cost accounting into one layer. Your application keeps a single address and a single credential; what changes behind that line stays behind it.

Four capabilities

Billing, availability, rules per business line, and data protection.

Billing

Spend broken down to the individual call

Every call records its business line, stage, attempt number and tool outcome. Monthly spend splits by team, product and vendor, and drills down into what it was made of.

Part of that invoice is money spent on retries.

  • Split by team, product line and vendor; drill down into retries, failed tool calls and discarded results.
  • One statement per month, locked on close, exportable for finance.
  • Enter your contracted vendor rates and vendor cost sits alongside platform fees on the same basis.
  • Alerts at half budget and near the cap. Balance is shown as days of runway.
What Tare sees
Illustrative — not a real customer invoice
On the invoice
What Tare sees
  • Output actually delivered61%
  • Retries after failure24%
  • Failed tool calls11%
  • Abandoned attempts4%
Availability

One vendor failing does not stop the work

A model name can map to several vendors, with primary, backup and weighting defined by you.

  • Timeouts or errors move traffic automatically, without waiting for an alert to be handled.
  • Recovery is confirmed by our own probe, not by your users retrying.
  • A persistently failing vendor is parked automatically and restored once it recovers.
  • If your own account fails, the platform account takes over and the model stays available.
  • A single call can name its own backup model, or point at, whitelist or block specific vendors — no platform change needed.
  • Every response can show how much of its time was the model answering, and how much was lost switching vendors.
  • Calls to your own vendor accounts can go through a node you host. Your application calls the node, the node calls the vendor, and nothing passes through us.
  • If we are unreachable, the node keeps working on the rules and balance it last received. Usage is reported once the connection is back.
Rules per business line

Rules enforced by configuration, not goodwill

One credential per business line, each with its own model permissions, budget and statement line.

  • Experiment credentials reach low-cost models only. Expensive models are outside their permissions.
  • Regulated work is pinned to approved vendors. Other credentials cannot reach them at all.
  • Spend separates by credential, and each credential carries a stated purpose. Every statement line maps to something real.
  • Alerts before the budget cap. Hard cut-off is a separate switch you turn on deliberately.
Data protection

Secrets and personal data are caught before they reach a vendor

Every call is checked. For each kind of content, you set what happens.

  • Checked: API keys, passwords, private keys, cloud credentials, card and ID numbers, emails, phone numbers. Names, places and organizations on request.
  • Four settings per kind: ignore, replace with a placeholder, fail the call, or forward it and send you an alert.
  • Set per account, tightened per credential where needed. Changes apply within seconds.
  • A node you host runs the same check on your machine. The content does not reach us.
  • Each hit is listed with the credential, the call, the kind of content and the action taken.
  • Alerts go to your own channels. One that did not arrive can be sent again from the console.
  • New kinds of leaks become new rules. We show what a rule would have caught on your traffic; you switch it on.

Two lines of configuration

No SDK, no agent, no change to how you call the API. Replace base_url and api_key.

Before
client = OpenAI(
    base_url="https://api.openai.com/v1",
    api_key=OPENAI_KEY,
)
After
client = OpenAI(
    base_url="https://tare.jamerly.ai/v1",
    api_key=TARE_KEY,
)
  • Model names, request and response shapes are unchanged. Streaming works as before.
  • Migrate one business line at a time. The rest is unaffected.
  • No extra instrumentation. The call already carries what the analysis needs.

Three ways to pay

By default we take no part in the token trade. We charge for governance, so we have no reason to want your usage to grow.

Your own vendor accounts
No charge

Your contracts, invoices and negotiated rates. Governance only, no transaction.

Your accounts, managed by us
Per token

A published rate per million tokens, independent of what you negotiated with your vendor.

Our vendor accounts
Token price

One balance, one invoice. No vendor accounts of your own to maintain.

Straight answers

How fine-grained is the governance?

One credential per business line, each with its own model permissions, budget and statement line. **There is deliberately no department or project layer above it** — a modelled org chart stops matching reality at the first reorg. Instead each credential carries a stated purpose, and that purpose travels with every figure it produces.

Is there a report for a quarterly vendor review?

Partly. Every park and restore is recorded with a timestamp, and calls and failures are counted per vendor. There is no availability percentage and no SLA report.

If a vendor drops mid-stream, does another finish the answer?

No. Vendor switching happens before output starts. Output already sent cannot be recalled, and two half-answers stitched together are harder to handle than a clear error.

Are the cost figures for my own accounts accurate?

Until you enter your contracted vendor rates they are a floor, and the page says so. They do not affect what we charge you — that is a published rate per token.

Does hitting the budget cap stop service?

It alerts by default. Hard cut-off is a separate switch you turn on deliberately.

Does the content of my calls pass through you?

Through our endpoint, yes. Through a node you host, calls to your own vendor accounts go straight from your network to the vendor. We receive the usage figures, not the content, and the check runs on your machine.

What happens when a call contains something sensitive?

What you set for that kind of content: nothing, a placeholder, a failed call, or a forwarded call plus an alert. Every hit is recorded. A missed alert can be sent again from the console.

Can a closed month change afterwards?

No. It locks on close. Corrections land in the open month and are visible to you.

Start with one business line

Move one across and leave it for a month. What comes back is a statement you can take into a meeting: who spent it, who it went to, and what it bought.

Talk through your setup