API portal
Console coming soon
Browse documentation

Developer docs

Start from a live reference page.

Phone layout stays search-first: pick a page here, then read it in a full single-column view.

Getting Started

OverviewPlatform map, live routes, auth model, and API quick links.QuickstartMake the first live request and inspect the response shape.AuthenticationBearer keys, scopes, rotation, and account-bound access.

Models

EvaEsi InstantByte-native chat with live evidence retrieval. The default model.Instant FastInstant with a shorter evidence gather, for latency-bound work.EvaEsi ReasoningEvidence-gather plus Elements synthesis for harder questions.Reasoning DeepA longer deliberation budget and the widest evidence sweep.Reasoning VerifyChecks claims against evidence and reports what it cannot support.Reasoning FastReasoning held to an interactive latency budget.ElementsCoding and Studio model. Routes directly without the evidence-gather pass.Elements FastSingle-file edits and short-range changes, tuned for latency.Elements CodebaseRepository-wide context for migrations and cross-file work.

API Reference

Responses APICanonical inference route for synchronous requests and output.Sources APIFetch evidence bundles and provenance for prior responses.Usage APICurrent period totals, history windows, and live limits envelopes.Models CatalogEvery model version, its canonical ID and its published rates.Batch APIDurable batches of up to 100 requests: submit, poll, collect.WebhooksSigned event deliveries: envelope, headers, and verification.StreamingServer sent events, frame shapes and handling failures mid stream.SchemaOpenAPI and JSON Schema documents, and generating a client from them.Error CodesOperational error contract, retry semantics, and failure classes.

Operations

Developer ConsoleThe console surfaces that open at launch, mapped honestly.Rungs & PricingBilling units, plan allowances, project budgets, and spend model.Rate LimitsHeaders, envelopes, throttling behavior, and budget enforcement.

Enterprise

EnterpriseReserved and dedicated deployments, isolation, data handling and controls.

Console

API keysIssue, scope and rotate API keys. Opens at launch.UsageRungs consumed, per model and per key. Opens at launch.BillingBudgets, invoices and payment. Opens at launch.SettingsOrganisation, members and security. Opens at launch.

Policies

Terms of ServiceOperational policy, legal terms, or support guidance.Privacy PolicyOperational policy, legal terms, or support guidance.Cookie PolicyOperational policy, legal terms, or support guidance.Community GuidelinesOperational policy, legal terms, or support guidance.Acceptable Use PolicyOperational policy, legal terms, or support guidance.Refund PolicyOperational policy, legal terms, or support guidance.Billing & TiersOperational policy, legal terms, or support guidance.HelpOperational policy, legal terms, or support guidance.API TermsOperational policy, legal terms, or support guidance.AI Use and Output SafetyOperational policy, legal terms, or support guidance.Accessibility StatementOperational policy, legal terms, or support guidance.SecurityOperational policy, legal terms, or support guidance.Service CommitmentsOperational policy, legal terms, or support guidance.SupportOperational policy, legal terms, or support guidance.ContactOperational policy, legal terms, or support guidance.

Streaming

Set stream to true and the response arrives as server sent events as the model produces it, instead of in one piece at the end.

Turning it on

Add "stream": true to the body of a request to POST /api/v1/responses. Everything else about the request stays the same. The response comes back as text/event-stream rather than JSON.

curl -N https://evaesi.dev/api/v1/responses \  -H "Authorization: Bearer $EVAESI_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "instant",    "prompt": "Explain rungs in two sentences.",    "stream": true  }'

Events

Two event types are sent. Each carries a JSON payload with the text so far in content and a done flag.

EventMeaning
response.output_text.deltaA new piece of the answer. Append it to what you have.
response.completedThe answer is finished. No further events follow.
event: response.output_text.deltadata: {"content":"Rungs are bytes","done":false}event: response.output_text.deltadata: {"content":". You are billed for","done":false}event: response.completeddata: {"content":"","done":true}

Treat response.completed as the only signal that the answer is whole. A connection that closes without it was cut short, and the text you have is partial.

Headers

Headers are sent when the stream opens, before any content exists. That means the output and total rung counts cannot be included: they are not known yet. You get the input cost and your rate limit position up front.

HeaderMeaning
x-evaesi-input-rungsRungs charged for the prompt.
x-evaesi-region-countRegions activated, when region reporting was requested.
x-ratelimit-limitRequests allowed in the current window.
x-ratelimit-remainingRequests left in the current window.

Errors mid stream

A request that fails before the stream opens returns a normal HTTP error, documented on the errors page. A failure after the stream has opened cannot change the status code, because 200 has already been sent. It arrives as an error frame instead.

event: error
data: {"error":{"code":"EVA_UNAVAILABLE","message":"Model unavailable."}}

This is the case most integrations forget. A stream that fails halfway looks like a successful request to any client that only checks the status code, so handle the error frame explicitly.

Billing

Streaming costs the same as not streaming. You are billed for the rungs you sent and the rungs that came back, whether they arrived in one piece or many. A stream you disconnect from early still bills for what the model produced before you left.

Example

The samples above open a stream and print each frame as it lands. For anything real you will want to parse the data: lines as JSON and accumulate content, stopping when you see response.completed.