Setup and current product limits

Make a bounded text request

1. Create an account and add a supported provider key in Dashboard → Provider Keys. The service stores an encrypted copy and decrypts it transiently for routing.
2. Create a ROUTEXOR API key in Dashboard → API Keys. Keep both types of key out of browser code, screenshots and version control.
3. Inspect GET https://api.routexor.com/v1/models. Choose a model ID whose provider your account can access. Catalog presence is not verified execution or a grant of provider access.
4. Set ROUTEXOR_API_KEY in your local environment, replace YOUR_ACCESSIBLE_MODEL_ID below, and send the request. This uses billable provider inference; choose your own test-spend ceiling first.

curl https://api.routexor.com/v1/chat/completions \
  -H "Authorization: Bearer $ROUTEXOR_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{"model":"YOUR_ACCESSIBLE_MODEL_ID","messages":[{"role":"user","content":"Say hello briefly."}],"max_tokens":64,"stream":false}'

Expect a Chat Completions-style response on success. A failure has an HTTP error status and safe error body; inspect the status before parsing a completion. Save your prior endpoint/configuration so you can reverse a migration.

Compatibility boundary

Try the offline migration preflight preview to check a redacted Chat Completions configuration against a small, dated contract subset. Analysis stays in your browser; no keys or requests are sent and no automatic cutover occurs.

• POST /v1/chat/completions and POST /v1/messages implement supported text and streaming paths. They do not promise every OpenAI- or Anthropic-format operation or parameter.
• The Responses API is not implemented. Strict JSON-schema output supports a bounded nonstreaming, pinned, text-only three-model subset. Other strict paths are explicitly refused; this is not complete OpenAI/OpenRouter compatibility.
• Assistant tool-call messages may have null or omitted content. Automatic cost, speed and quality selection preserves these histories and their tool results. Current local conformance tests exercise the authenticated router and adapters with synthetic provider responses; they are not certification of every model or client.
• Gemini vision translation supports base64 image data URLs for PNG, JPEG, WebP, HEIC and HEIF on normal and streaming Chat Completions. Remote image URLs, Files API references, Google tool round trips and thought-signature preservation are not verified migration paths. Unsupported or malformed Google image input is rejected without fetching an image URL from Routexor.
• Reasoning, cache accounting, tool execution and non-text modalities vary by model and adapter. A model capability label alone is not end-to-end conformance proof.
• Configured fallback requires compatible alternative models and usable provider keys. Once output starts, an upstream failure ends the stream with an error; the service cannot seamlessly replace already-delivered text.
• Claude Code and other agent clients must be checked at their actual version and workload. A base-URL change does not guarantee a complete coding-agent migration.

Routing controls and their limits

Eligible plans support task profiles for tagged requests and configured model/fallback preferences through the authenticated API. A complete self-service profile editor and an automatic routexor.config file loader are not yet available.

Per-call reservations allow for applicable catalog long-context, cache-write and peak-price rates. Completed normal and streaming calls use a captured admission-time price snapshot. Repeating the same acknowledged settlement does not subtract its reservation twice; uncertain or damaged accounting state requires review.

These are catalog-based provider-cost estimates, not fee-inclusive invoice ceilings. Token estimates, image billing, provider-added charges and upstream cancellation have limits. Keep provider-side limits in place. Reservation counters are not yet a restart-safe, automatically reconciled run ledger.

Request/key policy limits are not cumulative job budgets. Durable parent/child run tracking, shared reservations, automatic job pause and delivered budget alerts are not yet available. Bound retries, steps, fan-out and spending in your agent/provider accounts. In-flight requests can still incur provider charges.

Account-scoped routing health

Automatic selection and fallback ranking use only your account’s observations for its currently active provider keys. Rotating a key starts fresh evidence. Credential and quota errors do not change another customer’s routing.

Health uses a rolling five-minute window. A measured latency average requires at least five successful calls; a provider-failure rate requires at least five classified availability samples. Missing, stale or insufficient samples are unknown, not zero errors. Attempt duration includes streaming and consumer delays; it is not router overhead or time to first token.

Anonymous model listings describe catalog metadata only. With an active ROUTEXOR API key, GET /v1/models can return account_execution_observed for a successful call on your current key within 24 hours, plus routing_health. This is not capability certification, a global provider status or an uptime guarantee.

Ensemble costs and evidence

An eligible paid plan can request routexor/ensemble through Chat Completions. Members generate candidate answers and a judge synthesizes them. Multiple model calls can cost more and take longer than a single call; no quality improvement is guaranteed.

Provider charges still apply to every member and the judge. Only the adaptive-default judge is exempt from the Starter platform usage fee; custom ensembles and explicit member or judge overrides are metered normally.

Recorded aggregate usage includes successful members and judge usage. Failed or interrupted attempts may incur additional provider charges. Agreement with a judge is not verified task correctness, and a modeled cost difference is not demonstrated net savings.

Plans and billing

Provider charges are separate from ROUTEXOR platform fees. Starter is usage-metered; Pro is a flat subscription. Provider and service limits still apply.

Recorded provider cost is an estimate, not a reconciled invoice. It excludes ROUTEXOR platform fees and may differ from provider billing. Verified net savings are not available. Checkout availability and payment status must be confirmed. Billing reconciliation is not represented as complete by this guide.

Security, privacy and planned capabilities

Provider keys are encrypted at rest, not customer-only decryptable. Session revocation and safe-error controls are implemented; this is not a claim of independent security certification or absence of security gaps. Optional evaluation workflows can retain supplied messages; see the privacy policy.

CORTEX memory integration, outcome-verified policy optimization and durable agent-run controls remain planned. Enterprise infrastructure, support and contractual commitments require an agreed scope; no certification, region guarantee or SLA is implied by a plan label.