Providers
30+ providers. Your keys, or ours.
Bring your own API keys and pay providers directly, or use Platform Access and skip the provider account.
How access works
v17 live
THE AI CONTROL PLANE
One API. Every model. Your rules.
Route AI requests across providers, fall back when one breaks, cap spend per key, and trace every call in real time.
providers
30+
bring your own key to any of them
markup on your keys
0%
your provider bills you directly
routing strategies
4
plus automatic fallback
official SDKs
3+ CLI
JavaScript · Python · PHP
The two-minute tour
From provider keys to production traffic.
Connect a provider, build a route, ship one API key, and watch the requests arrive.
Routing, live
See every routing decision.
Each request is checked against its key, its spending limit, and its route. LatentKit ranks the models by your strategy, skips providers that are failing, and falls back on its own when one does.
- 1Check the key. Credential valid, key switched on, monthly limit not reached.
- 2Rank the route. Drop models that can't do the job, order the rest by your strategy.
- 3Call and recover. On a rate limit, error, or timeout, the next model takes over.
Purposes
Name the job. LatentKit picks the model.
Tag the models in a route with purposes like support-reply or summarize. Your app sends the label, and the route decides which model answers. Swap the model in the console and the code never changes.
- Its own strategy per purpose: priority, cost, availability, or balanced weights.
- Safe by default: an unknown label just uses the route's normal order.
- Visible in usage: spend and requests break down by purpose.
const reply = await client.chat.create({
purpose: "support-reply",
messages,
});
production-chat
v17 live
support-replyPriority2 models- 1claude-sonnet-4-5Anthropic · BYOKanswers
- 2gpt-4.1OpenAI · Platformfallback
Then the rest of the route, in order.
lk_applied_purpose "support-reply" model claude-sonnet-4-5
const summary = await client.chat.create({
purpose: "summarize",
messages,
});
production-chat
v17 live
summarizeCost2 models- 1deepseek-chatDeepSeek · Platformcheapest
- 2gemini-2.5-flashGoogle · BYOKfallback
Then the rest of the route, in order.
lk_applied_purpose "summarize" model deepseek-chat
const fields = await client.chat.create({
purpose: "extract",
messages,
response_format: { type: "json_object" },
});
production-chat
v17 live
extractBalanced2 models- 1gpt-4.1-miniOpenAI · Platform70%
- 2mistral-smallMistral · BYOK30%
Weights split the first attempt. Every other model stays a fallback.
lk_applied_purpose "extract" model gpt-4.1-mini
One console
Connect it. Control it. Test it. See it.
Everything between your app and the model lives in one place, so the team that ships the feature can also run it.
API keys
One key per app, with a limit.
Shown once, capped per month, rotated in a click, switched off instantly.
Key controlsPlayground
Try the route before you ship.
Real requests through a key's live route, with the model, latency, and cost on every reply.
See the PlaygroundRequest logs
Every call, with the receipt.
Which model answered, what it tried first, why it fell back, and what it cost.
What gets recordedDevelopers
Write the call once.
Requests never name a model, so switching providers, adding fallbacks, or changing cost rules is a console change, not a deploy. Streams, tool calls, and structured output come back in one shape, whichever provider answers.
- JavaScript
npm install @latentkit/sdk - Python
pip install latentkit - PHP
composer require latentkit/latentkit-php - CLI
npm install -g @latentkit/cli
import { LatentKit } from "@latentkit/sdk";
const client = new LatentKit(); // reads LATENTKIT_API_KEY
const reply = await client.chat.create({
purpose: "support-reply",
messages: [{ role: "user", content: "Where is my order?" }],
});
console.log(reply.content, reply.model);from latentkit import LatentKit
with LatentKit() as client: # reads LATENTKIT_API_KEY
reply = client.chat.create(
purpose="support-reply",
messages=[{"role": "user", "content": "Where is my order?"}],
)
print(reply["content"], reply["model"])use LatentKit\LatentKit;
$client = new LatentKit(); // reads LATENTKIT_API_KEY
$reply = $client->chat->create(
purpose: 'support-reply',
messages: [['role' => 'user', 'content' => 'Where is my order?']],
);
echo $reply['content'];curl https://ai.latentkit.com/v1/chat \
-H "Authorization: Bearer $LATENTKIT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"purpose": "support-reply",
"messages": [{ "role": "user", "content": "Where is my order?" }]
}'
// response
{ "provider": "anthropic", "model": "claude-sonnet-4-5", "cost_usd": 0.0021, … }
Why a gateway
Stop rebuilding the same AI plumbing.
Calling providers directly works until the second provider, the first outage, or the first surprise invoice.
When you need to
Calling providers directly
With LatentKit
Switch models or providers
Change code, redeploy every service
Publish a new route version
Survive an outage or rate limit
Write and maintain retry logic
Automatic fallback to the next model
Keep provider keys safe
Provider keys in every app's config
One LatentKit key per app; provider keys stay in the vault
Control spend
One limit for the whole provider account
A monthly limit on each key
Explain a bad answer
Search your own logs
A record of every attempt, model, latency, and cost
Test a change
Write a script
Run it in the Playground on the live route
For companies
Run AI across the company, under your brand.
Company workspaces add team roles, your own dashboard domain and branding, scheduled rollouts, an audit trail, and limits set for you.
Pricing
Start free. Pay for what you route.
Your own provider keys carry no LatentKit markup. Platform Access is provider cost plus 5%, from prepaid credits.
Ship AI traffic you can trust.
Start free with your own provider keys. No credit card.