Wrengle AI and credits

Wrengle AI is the prepaid alternative to using an OpenAI or Anthropic key of your own. A clean install starts the assistant on Wrengle AI · Fast, then shows the destination, storage, and charging disclosure before its first managed request. Existing explicit provider and Auto choices remain unchanged when upgrading. You can switch routes at any time.

Creating an account or signing in does not include free Wrengle AI credit. There is no recurring allowance, trial balance, subscription, automatic top-up, or overage plan. A request starts only when the available balance can cover its reservation. Buying credits is optional; using your own provider key remains available separately.

Tier mappings and rates

All three Wrengle tiers have published mappings and rates. A deployment exposes only the subset named in its WRENGLE_AI_ENABLED_TIERS allow-list; an absent or empty allow-list exposes none of them. The app therefore shows the tiers that the signed-in account endpoint reports, rather than assuming every priced tier is currently available.

Rates below are per 1 million tokens. Input is the provider's total input count; the lower cached and cache-write rates replace, rather than add to, the ordinary input rate for those token buckets.

One credit represents US$0.01 of underlying model usage. The corresponding OpenAI rates per 1 million tokens are US$0.20 / US$0.02 / US$0.25 / US$1.20 for Fast, US$2 / US$0.20 / US$2.50 / US$12 for Balanced, and US$4 / US$0.40 / US$5 / US$20 for Deep, in the same input / cached-input / cache-write-input / output order. Credit-pack purchase prices are separate: the difference between a pack price and its underlying model-usage value is Wrengle's margin. Each request rounds its charge up to the nearest millionth of a credit.

When total input exceeds 272,000 tokens, the whole request uses twice the listed rate for every input category and one-and-a-half times the listed output rate. Reasoning tokens are recorded and are already included in the provider's output count, so they are not added a second time.

Current request limits

Wrengle AI is text-only in this release. Image attachments are supported on the compatible direct OpenAI and Anthropic routes, not on the prepaid route. Wrengle AI also accepts only the fields and output ceiling the gateway has reviewed; choosing a tier never turns the account route into a general OpenAI proxy.

Auto still considers only configured provider keys and never selects Wrengle AI. The clean-install Fast route is a concrete saved default, not an Auto result. Wrengle AI does not fall back to your own provider key, another tier, or a local model, and the gateway does not retry a request through another destination. Provider replies are limited to 8 MiB in total, including streamed replies. If a provider exceeds that boundary, Wrengle cancels the stream and applies the interrupted-request estimate described below.

Signing out on desktop, or replacing the signed-in account, cancels that account's in-flight managed work. Late assistant tokens and results are not shown, and managed note, workspace-answer, workflow, voice-analysis, meeting report, and plugin-draft writes cannot cross into the replacement session. Start the work again after switching accounts. A provider request that was already accepted can still have incurred usage; if its final usage block is no longer available, the interrupted-request rule below applies.

What the ledger records

For each provider dispatch, Wrengle records:

• the account, feature, Wrengle tier, and OpenAI model;
• the Wrengle request ID and OpenAI request ID when OpenAI supplies one;
• provider-reported input, cached-input, cache-write-input, output, and reasoning token buckets;
• the credit charge, price-version number, status, and time.

It does not record the prompt, reply, note text, transcript text, attached context, or tool arguments in the credit ledger. That storage boundary does not mean managed content stays local: the request and response still follow the transient route described below.

A completed provider reply normally has a Provider reported status. Expand Audit details on its dashboard row to see both request IDs, price version, and every reported token bucket separately. If a stream ends or the caller disconnects before OpenAI supplies the final usage block, Wrengle cannot honestly state a token count. It charges the request's reserved ceiling, marks the row Estimated charge, and shows Token counts unavailable instead of claiming that zero tokens were used.

The account dashboard always shows the ledger when the account exists. Checkout is a separate deployment capability: when purchases are unavailable, balances, holds, and usage remain visible but Buy and receipts controls do not appear.

Content path and privacy

Prompts and replies transit Wrengle and OpenAI for generation. It is not a local-only operation: bounded prompt and context are uploaded to Wrengle, sent on to OpenAI, and the reply returns through Wrengle. That content is not stored in the account ledger or application logs by Wrengle; the ledger keeps only the usage metadata listed above. OpenAI receives the content and applies its own service handling, logging, and retention terms.

See AI and privacy for the feature-by-feature content bounds, account-token handling, provider exposure, and local data that never enters this route.

Deployment states

The hosted gateway and credit checkout are independently configured:

• If the gateway is off, Wrengle AI generation is unavailable and no request can reach OpenAI through Wrengle.
• If checkout is off, existing balances and usage still appear, but the account cannot start a purchase or open the Polar receipts portal.
• Turning checkout on does not turn generation on, and turning generation on does not create or grant credit.

Operators configure the gateway with WRENGLE_AI_GATEWAY_ENABLED=1, a server-side OPENAI_API_KEY, and the comma-separated WRENGLE_AI_ENABLED_TIERS allow-list. Missing gateway configuration or an empty tier allow-list fails closed. Polar access, webhook, and credit-pack variables control checkout separately. Neither secret is shipped in the desktop app or browser bundle.

Optional safety overrides accept positive whole numbers only. WRENGLE_AI_MAX_OPEN_OPERATIONS defaults to 4 and is capped at 32 per account; WRENGLE_AI_OPERATION_TTL_MS defaults to 30 minutes and is capped at 24 hours; and WRENGLE_AI_UPSTREAM_TIMEOUT_MS defaults to 120 seconds and is capped at 30 minutes. Invalid or larger values use the safe defaults. Explicit multi-call operations must also fit their timeout budget and five-minute close margin inside 24 hours.