Closed beta · a RiKuWe GmbH project

AI costs are hard to predict.

If you run LLM APIs in production, you pay per token. At the end of the month there is a total on the invoice that can only be assigned to individual teams, models or use cases with considerable effort. TOKENWERK is the gateway in between.

Request beta access Read the docs
2 API formats OPENAI · ANTHROPIC
EU run and hosted in the EU GDPR-COMPLIANT USE POSSIBLE
0 prompts stored METADATA ONLY, NO CONTENT
Curated models by purpose TESTED AND MAINTAINED BY US
Where we are today CLOSED BETA
  • One way into the models of several providers. One base URL, the SDKs you already use. live
  • Our own model names across providers. One name, routes behind it by priority; if a provider fails, the next takes over. live
  • Usage and cost per request. Billed by input, cached input and output tokens, visible by key, team and model. live
  • Budgets for the day, the week and the month. Per key or team, with a warning email, after which requests are declined instead of adding up. live
  • Compression of long contexts. Redundant context is dropped, so fewer tokens reach the model. coming soon
  • Automatic model choice per request. The gateway decides which model a request needs. in progress
Failover · try it EXAMPLE

Interactive example: requests from your app go through TOKENWERK to three routes in order of priority. Switch a provider off with its button and TOKENWERK tries every request with the next one; only when none is reachable does your app get an error.

Your app POST /v1/chat/completions
TOKENWERK Routes by priority

Pull the plug on a provider: just click it.

The gateway

One interface, six jobs.

Your applications always call the same base URL. Behind it, TOKENWERK picks model and provider, records the usage and enforces the budgets.

01 · ACCESS

A compatible API

TOKENWERK speaks the OpenAI and Anthropic APIs under one domain. You keep your SDK and your tooling, what changes is the base URL and the key.

api.tokenwerk.ai

02 · MODELS

Curated models

A maintained selection under our own names: you choose by use case, not by trends. Which provider answers behind a name is our job. Models and prices

priority + fallback

03 · COST

Usage in plain sight

Usage is recorded per request: input, cached and output tokens, duration and price, at the model price when the request started. Broken down by key, team and model.

per request

04 · BUDGETS

Limits, not surprises

Keys belong to an organisation, optionally to a team, and can carry an expiry date and budgets for the day, the week and the month, with budget warnings by email. Once a budget is reached, requests are declined.

day · week · month

05 · COMPRESSION

Shorter contexts

Removes redundant context from your requests before they reach the model: fewer tokens per call, lower cost. The longer a conversation grows, the more you save.

coming soon

06 · ROUTING

Automatic model choice

Not every request needs the largest model. The gateway picks the fitting one instead of you naming it.

in progress

Integration

Swap the base URL. That is it.

TOKENWERK sits as a gateway between your application and the models. No SDK change, no new abstraction: you set the base URL and the key, the rest of your code stays as it is.

  • Operated and hosted in the EU. That makes GDPR-compliant use possible. Whether your specific use is compliant depends on your own case.
  • Streaming, tool calling, structured outputs and image inputs work as usual. Your request goes to the provider unchanged, unless you switch the optimisation on.
  • No model name maintenance: you use a TOKENWERK model and we make the exact choice for you.

Read the docs · Models and prices

from openai import OpenAI

# Before
client = OpenAI(
  api_key="sk-..."
)

# After, that is all
client = OpenAI(
  api_key="tw-...",
  base_url="https://api.tokenwerk.ai/v1"
)

# Your code stays unchanged
response = client.chat.completions.create(
  model="MODEL",
  messages=[{"role": "user", "content": prompt}]
)

Same with the Anthropic SDK: x-api-key: tw-…

Context compression coming soon

Preview, coming soon: TOKENWERK removes redundant context from long conversations before a request reaches the model. The longer the conversation, the bigger the difference between what your app sends and what reaches the model. Fewer tokens at the model mean lower costs.

Your app 8 messages in the conversation
TOKENWERK Compression
Model gets the compact context

Redundant context is dropped before the request reaches the model: fewer tokens, lower costs. The longer the conversation, the more you save.

TOKENWERK is in closed beta.

We work with a limited number of teams running LLM APIs in production and build out the layers together with them. There is no open sign-up yet: we hand out access individually and reply in person.

Access on request · run in the EU · usually an answer within 24 hours

Request beta access

TOKENWERK is a project of RiKuWe GmbH from Styria, Austria. You talk directly to the people working on it.

More about the company behind it at rikuwe.com.

RiKuWe GmbH
Entschendorf 61
8321 St. Margarethen an der Raab
Austria