What can mTokOps track?
Not every provider shares the same information. This guide separates what is built from possible future connections. Listing a provider does not mean it is connected, available to add, or verified against its bills.
No provider is live-reconciled yet. Current integrations still have known accounting or coverage gaps. “Built” means code exists, not that its totals are billing-ready. “Not built” means a researched option, not an available feature or a delivery promise.
Collection methods, in plain English
- Account reports
- Import what the provider records. Can include earlier usage, within its reporting limits. Usually needs admin permission.
- Cloud metrics + billing
- Combine activity measurements with a separate cloud bill. They may arrive at different times and offer different detail.
- Capture new requests
- Add tracking to your software or a gateway. Only requests passing through it are captured; it does not recover earlier traffic.
- Specialized collectors
- Read local runs or prediction metrics when standard token reports do not fit.
- Report files
- A possible fallback when a provider offers usable exports. File import is not built here and would not be automatic monitoring.
Current focus — existing integrations
OpenAI, Anthropic, Gemini and Abacus.AI remain the implementation priority. Their limitations are shown individually below.
OpenAI
Built · corrections needed
- How we would collect it
- Import account reports
- Usage detail
- Daily tokens and requests, with project, model and key detail.
- Money
- Provider-reported account/project costs; key-level costs are not collected.
- What to know
- Admin reporting access required. Audio counting and import integrity need correction; cached-token detail is not retained. Totals are not yet reconciled against a live provider report.
Anthropic
Built · corrections needed
- How we would collect it
- Import account reports
- Usage detail
- Daily Messages tokens by workspace, model and key.
- Money
- Provider-reported costs; the reporting endpoint excludes Priority Tier costs.
- What to know
- Admin reporting access required. Cache breakdowns and tool counts are not retained; import integrity needs correction. Totals are not yet reconciled against a live provider report.
Gemini / Google AI Studio
Built · limited coverage
- How we would collect it
- Read Google project metrics
- Usage detail
- Partial request and quota metrics, not complete input/output or thinking-token usage.
- Money
- Incomplete estimates, not actual billing.
- What to know
- Google monitoring permissions required. Current output-token zeros do not prove no output usage. Accurate token and billing collection needs redesign; this is not Vertex AI coverage.
Abacus.AI / RouteLLM
Reporting source unverified
- How we would collect it
- Separately configured reporting feed
- Usage detail
- Depends on a usable reporting feed; account-wide history has not been established.
- Money
- Reliable billed-cost coverage has not been established.
- What to know
- A valid model-access key does not grant historical reporting. The custom-feed adapter exists, but plug-and-play monitoring is not verified.
Wider coverage — researched, not built
Every other provider and billing route reviewed so far is listed below. These are lower-priority expansion options; no live access or collection has been verified.
OpenRouter
Not built
- How we would collect it
- Import account activity reports
- Usage detail
- Candidate: activity for requests routed through OpenRouter.
- Money
- Reported activity costs, subject to billing-route reconciliation.
- What to know
- Management key required; the reviewed activity endpoint covers the last 30 completed UTC days. Not traffic sent directly to other providers.
xAI
Not built
- How we would collect it
- Import team usage reports
- Usage detail
- Candidate: team usage through the Management API.
- Money
- Cost fields and scope need validation before presenting billed totals.
- What to know
- Management access required; a regular inference key alone is not proof of reporting access.
Mistral
Not built
- How we would collect it
- Import admin usage reports
- Usage detail
- Candidate: organization usage through the Admin API.
- Money
- Billed-cost availability and detail still need validation.
- What to know
- Admin reporting permissions required; history and attribution must be checked during integration.
Fireworks
Not built
- How we would collect it
- Import usage and cost reports
- Usage detail
- Candidate: account usage reporting.
- Money
- A usage-cost query endpoint exists; coverage still needs validation.
- What to know
- Account-wide cost queries require account-admin access. Not every usage endpoint provides the same cost detail.
Together AI
Not built · access-dependent
- How we would collect it
- Account reports or capture new requests
- Usage detail
- Candidate: billing/usage reports where enabled, or token counts returned with requests.
- Money
- Reporting access may provide costs; request tokens alone do not establish the bill.
- What to know
- The reviewed reporting API is a per-organization beta. Request capture only sees traffic sent through the integration.
Azure OpenAI
Not built
- How we would collect it
- Combine cloud metrics and billing
- Usage detail
- Candidate: Azure resource/deployment metrics.
- Money
- Azure billing data would be collected separately from usage metrics.
- What to know
- Azure account permissions required. Reporting delays and model/token breakdowns vary; direct OpenAI reporting does not cover this route.
Amazon Bedrock
Not built
- How we would collect it
- Combine cloud metrics and billing
- Usage detail
- Candidate: CloudWatch usage metrics for supported models and request types.
- Money
- AWS billing data would be collected separately.
- What to know
- AWS permissions required. Model and request-type coverage varies; Anthropic direct reports do not cover Bedrock usage.
Google Vertex AI
Not built
- How we would collect it
- Combine cloud metrics and billing
- Usage detail
- Candidate: Vertex project/model metrics and supported response usage.
- Money
- Google Cloud billing exports would provide a separate cost source.
- What to know
- Google Cloud permissions required. Separate from AI Studio; publisher models billed through Vertex belong to this route.
Groq
Not built · access-dependent
- How we would collect it
- Capture new requests; enterprise metrics where available
- Usage detail
- Candidate: returned token counts; enterprise accounts may expose Prometheus metrics.
- Money
- Request-based costs would be estimates unless a billed-cost source is added.
- What to know
- Enterprise metrics are not available to every account. Request capture does not recover account-wide history.
DeepSeek
Not built
- How we would collect it
- Capture new requests
- Usage detail
- Candidate: usage fields returned with individual requests.
- Money
- Rate-based estimates; the balance endpoint is not an itemized usage or cost report.
- What to know
- Requires changes to the software making requests. Only captured traffic is covered; historical reporting is not established.
Cohere
Not built
- How we would collect it
- Capture new requests
- Usage detail
- Candidate: tokens and billed-unit fields returned with requests.
- Money
- Billed units are not dollar costs; a rate-based calculation would be an estimate.
- What to know
- Requires request integration. Endpoint fields vary, and prior account-wide usage is not recovered.
Perplexity
Not built
- How we would collect it
- Capture new requests
- Usage detail
- Candidate: returned token and usage fields.
- Money
- Response-reported costs where present; otherwise unavailable or explicitly estimated.
- What to know
- Requires request integration. Cost fields depend on the endpoint; captured requests are not a full historical account report.
Cerebras
Not built
- How we would collect it
- Capture new requests
- Usage detail
- Candidate: token counts returned with supported requests.
- Money
- Rate-based estimates unless a billed-cost source is established.
- What to know
- Requires request integration; only captured traffic is covered, not earlier account history.
Hugging Face
Not built
- How we would collect it
- Route-specific request and billing collection
- Usage detail
- Candidate: usage returned by the selected inference route.
- Money
- Depends on whether Hugging Face or a separately keyed provider bills the request.
- What to know
- Routed and bring-your-own-key usage must stay separate. No single uniform token/cost report has been validated for this app.
Replicate
Not built
- How we would collect it
- Collect prediction metrics
- Usage detail
- Candidate: prediction runs and execution metrics, rather than a universal token total.
- Money
- Runtime metrics alone do not establish billed costs; pricing depends on the model.
- What to know
- Model outputs and billing units differ. A specialized collector and billing reconciliation are needed.
Ollama
Not built
- How we would collect it
- Capture local model responses
- Usage detail
- Candidate: prompt/evaluation counts and timing returned by local runs.
- Money
- No provider API bill for local execution; hardware, hosting and electricity costs are separate.
- What to know
- Requires access to the calling software or local service. Counts are not automatically comparable to hosted-provider billing.
What this means for your companies
Use dedicated projects, workspaces or keys to separate each company’s usage where reporting supports it. Key-level tokens do not automatically give key-level dollars. Shared costs that cannot be assigned confidently should remain Unallocated.
The billing route matters more than the model name: Claude through Anthropic, Bedrock and Vertex are different sources. A gateway sees only traffic routed through it. Do not add gateway or request estimates to the provider’s bill for the same calls.
Unknown is not zero. Missing output, cache, reasoning or cost fields are coverage gaps, not proof of no usage. Check the last successful import and the provider’s reporting delay; daily reports are not a live request feed.
Promptfoo is a testing tool, not another provider. Its calls belong to the provider and company that pay for them. A separate test-results import is deferred; it must not duplicate provider costs.