# Provider support research This document records providers considered for usage-bar support. Recheck blocked findings when a provider publishes a new quota, balance, or billing API. ## Qwen Token Plan **Status:** Blocked — no account-usage surface found as of 2026-08-13. Pi 0.84.1 supports these provider IDs: - `qwen-token-plan` - `qwen-token-plan-individual` - `qwen-token-plan-cn` The Individual provider shares the international inference endpoint and `QWEN_TOKEN_PLAN_API_KEY` with the existing international provider, but exposes a narrower model catalog. Research found no documented API-key-authenticated endpoint that reports consumed or remaining quota, percentages, or reset times. Independent probing of eight likely usage/quota paths on `token-plan.ap-southeast-1.maas.aliyuncs.com` returned `404`, and successful inference responses did not include quota or rate-limit headers. Qwen's client can recognize an exhausted-quota error, but that provides no usage percentage or reset timestamp. Do not estimate account-wide quota from requests observed by this extension: other clients and sessions would make that value incomplete and misleading. **Recheck when:** Alibaba Model Studio documents a Token Plan usage endpoint, adds quota headers to inference responses, or exposes a key-authenticated portal API. ## Baseten **Status:** Supported as of 2026-08-28. Pi's `baseten` provider resolves `BASETEN_API_KEY`. Baseten documents `GET https://api.baseten.co/v1/billing/usage_summary`, authenticated with `Authorization: Bearer $BASETEN_API_KEY`, with required UTC `start_date` and `end_date` parameters (maximum range: 31 days). The response reports `credits_used` for dedicated serving, training, and Model APIs. The extension queries the current UTC calendar month and reports the aggregate Credits used; it intentionally does not manufacture a remaining balance or quota percentage. **Recheck when:** Baseten publishes a first-party balance, budget, or quota API that permits a more complete account indicator. ## Vercel AI Gateway **Status:** Supported as of 2026-09-24. Pi's `vercel-ai-gateway` provider resolves `AI_GATEWAY_API_KEY`. Vercel documents `GET https://ai-gateway.vercel.sh/v1/credits`, authenticated with the same API key as a bearer token. The response reports the team's remaining AI Gateway Credits balance and total lifetime spend in USD. **Recheck when:** Vercel adds defined-period spend or budget fields that provide more actionable detail than lifetime spend. ## Fireworks **Status:** Supported as of 2026-09-24. Fireworks documents bearer-key-authenticated account discovery through `GET https://api.fireworks.ai/v1/accounts` and account-wide rated costs through `GET /v1/accounts/{account_id}/billing/summary`. Fireworks states that any key on the target account can use the API. A live redacted test with an ordinary user API key resolved through Pi's `fireworks` provider returned HTTP 200 for account discovery, billing summary, quotas, and the self-scoped usage-cost query. It exposed one accessible account; no credential, account identifier, or raw response was retained. The extension reports rated spend for the current UTC calendar month from the billing summary. It does not use the `monthly-spend-usd` quota as a budget percentage because validation observed a `9,999,999` default that appears to mean effectively unlimited rather than an actionable budget. Exactly one accessible account is selected automatically. When a key can access multiple accounts, `PI_FIREWORKS_ACCOUNT_ID` must match one of the discovered account IDs; the extension does not guess or aggregate accounts. **Recheck when:** Fireworks documents an account credit balance or a reliable distinction between explicit spend budgets and effectively unlimited quota values. ## xAI **Status:** Supported as of 2026-09-24. xAI documents its billing APIs at `https://management-api.x.ai`, authenticated with a Management Key that is deliberately separate from the normal inference API key or subscription OAuth credential used by Pi's `xai` provider. The extension registers the model-less `pi-usage-bars-xai-management` provider so `/login` can prompt for and store a read-only Management Key through Pi's credential system. `XAI_MANAGEMENT_KEY` remains an optional environment fallback. A live redacted test with a read-only, team-scoped key returned HTTP 200 for key validation, prepaid balance, current invoice preview, and postpaid spending limits. The extension validates `SCOPE_TEAM`, obtains the team ID internally, and reports the prepaid accounting total as a positive available balance plus `coreInvoice.totalWithCorr` as current billing-cycle spend. It never logs the key, team ID, ACL contents, or raw response. Organization-scoped keys are rejected rather than guessing or aggregating teams. **Recheck when:** xAI adds team discovery for organization-scoped Management Keys or exposes billing through the ordinary inference credential. ## Meta Muse **Status:** Blocked — no documented account-usage API found as of 2026-09-24. Pi 0.86.1 added the `meta` provider with Meta Muse subscription OAuth and `META_API_KEY` authentication. No documented first-party endpoint was found for remaining subscription quota, reset timing, account balance, or spend. Do not infer quota from local requests or use undocumented console endpoints. ## Acceptance criteria for a new provider A provider is suitable when a documented first-party API or response header supplies at least one meaningful account-level value: - quota consumed or remaining, ideally with a reset time; - account balance or credits; or - spend for a defined billing period. Authentication must work through Pi's provider credential API. The extension must not import browser cookies, scrape console pages, estimate account-wide usage from local traffic, or send credentials to third-party services.