Start free
PlatformHow LatentKit works

One gateway between your app and every model.

Your app sends one request with one key. LatentKit checks the key, applies the route you published, picks a model, falls back if it fails, and records what happened.

Start free → Core concepts ↗

Your app

POST /v1/chat Authorization: Bearer lk-… purpose: "support-reply"
  1. 1AuthenticateCredential valid, key on, monthly limit not spent.
  2. 2Load the routeThe key's published route, purposes, and depth.
  3. 3FilterDrop models that can't do the job or are failing.
  4. 4RankOrder the rest by strategy and purpose.
  5. 5Call, fall backErrors, limits, and timeouts go to the next model.
  6. 6RecordModel, attempts, latency, tokens, cost.

30+ providers

Routing

Routes decide. Your code doesn't have to.

A route is an ordered set of models and the rules for choosing between them. Point any number of API keys at it. Publish a change and every one of them follows, with no deploy.

production-chat v17 live 3 API keys Up to date
Strategy
Default depth
Models4
  1. 1 claude-sonnet-4-5Anthropic · BYOK $3.00 / $15.00
  2. 2 gpt-4.1OpenAI · Platform $2.00 / $8.00
  3. 3 gemini-2.5-flashGoogle · BYOK $0.30 / $2.50
  4. 4 deepseek-chatDeepSeek · Platform $0.27 / $1.10

Your order. Each model below only runs if the ones above it fail.

Fallback, built in

Rate limits, provider errors, timeouts, and empty answers move the request to the next model, up to three times by default. Providers that keep failing are skipped until they recover.

Only models that can do the job

Images, tool calls, JSON output, audio: a model that can't handle the request is never tried.

Fast, balanced, or thinking

A request can ask for speed or depth. The route reorders its models and sets each one's reasoning effort to match.

Versions you can undo

Every publish is a numbered version. Discard a draft or roll back; Company plans can schedule a publish with a canary.

Purposes

Different jobs, different models. Same route.

Purposes group a route's models by job. Each one has its own models, order, and strategy, and your app chooses it by name: purpose: "summarize".

  • Rename nothing in code when you change which model handles a job.
  • Balanced weights split a purpose's traffic, from 1 to 100 per model.
  • Graceful: an unknown purpose uses the route's normal order, and the response says which purpose applied.
Read the purposes guide

Providers

30+ providers. Your keys, or ours.

Connect a provider with your own API key and pay it directly, or turn on Platform Access and pay through LatentKit credits. One route can mix both.

BYOK

Your API key

You already have a provider account.

  • Billed byyour provider, directly
  • LatentKit markup0%
  • Providersall 30+, plus your own OpenAI-compatible endpoints on paid plans

Platform Access

LatentKit's access

You want to start without a provider account.

  • Billed fromprepaid LatentKit credits
  • Priceprovider cost + 5%
  • ProvidersOpenAI, Anthropic, Gemini, Mistral, DeepSeek, xAI, OpenRouter

Connect with your own key

OpenAI · Anthropic · Google Gemini · Mistral · DeepSeek · xAI · OpenRouter · Groq · Together AI · Fireworks · Perplexity · Cohere · Cerebras · DeepInfra · Replicate · fal · AWS Bedrock · Azure AI Foundry · Google Vertex AI · Alibaba · ByteDance · Moonshot · MiniMax · Z.ai · SiliconFlow · Scaleway · Jina · AssemblyAI · Deepgram · ElevenLabs · Google Speech

  • Provider keys go into an encrypted vault and are never shown again.
  • New connections are tested with the provider before they are used.
  • Turn individual models on or off for each connection.

API keys

One key per app. Limits you can see.

Give every app and environment its own LatentKit key, pointed at a route. Provider credentials never leave LatentKit, and a leaked key is one click from dead.

  • Shown once, then stored only as a hash.
  • A monthly spending limit and an expiry date on each credential.
  • Rotate in place: the old secret stops working at once.
  • Switch off a key, or a whole route, in one click.
  • Optional response cache per key, exact or semantic.

Request logs

Know what happened to every request.

Every call is recorded with its route version, each model it tried, why it fell back, and the latency, tokens, and cost of the answer.

Find the request

Filter by provider, model, status, and time. Sort by cost or tokens.

See the trend

Requests, spend, tokens, errors, P50 and P95 latency, and your top keys by spend.

Break it down

Each key's spend by model, by provider, and by purpose. Your keys and Platform Access, apart or together.

Playground

Try the route before you ship the code.

Pick an API key and send a message. It runs through that key's live route, so the model, latency, and cost under each reply are the real ones.

See the Playground

Security

Built to hold your provider keys.

  • Provider keys are encrypted at rest, never logged, and can't be read back.
  • LatentKit API keys are stored only as hashes.
  • A strict zero-retention mode keeps request metadata only.
Security posture

Route your first request today.

Connect a provider, publish a route, and get a key in the guided setup.

Start free → For developers