Set up in minutes, not meetings.

1 · Put the numbers in your chat free · open source

Works with a Wendl gateway or any LiteLLM-compatible proxy you already run. No agent calls, no tokens burned.

OpenClaw — two commands:

$ openclaw plugins install git:github.com/wendl-ai/openclaw-wendl@main
$ openclaw plugins enable wendl

Then add your proxy key to openclaw.json and restart:

"plugins": { "entries": { "wendl": { "config": {
  "litellmBaseUrl": "http://localhost:4000",
  "litellmMasterKey": "sk-…"
} } } }

Type /limits (caps & spend), /stats (savings), /wendl (managed setup). Full reference in the plugin repo.

2 · See what you'd save — before installing anything

We replay a slice of your real usage against today's model catalog and send back a breakdown: what you spend now, what routing would have cost, where the delta comes from.

What to send

1. Easiest: your provider's usage export — the CSV/JSON from your OpenAI or Anthropic billing/usage page.
2. Best: an OpenTelemetry trace export from your agent stack (GenAI semconv — most frameworks emit it). Latency and error rates make the breakdown sharper.

Email it to [email protected] — you'll get the report back as a private page. See a sample →

3 · Turn on routing

The Wendl gateway is a single Node process — no Docker, no Postgres. It speaks both OpenAI and Anthropic APIs, so agent harnesses point at it with two env vars.

Self-hosted early access

Three commands from a checkout:

$ make bootstrap-lite   # seeds team + ops budget keys, prints them
$ make gateway          # serves :4000 — bound to localhost by default
$ make rotate-key ALIAS=team   # if a key ever leaks: old dies, budget survives

Point your harness at it (any Claude-SDK harness, incl. nanoclaw and Claude Code):

ANTHROPIC_BASE_URL=http://localhost:4000
ANTHROPIC_AUTH_TOKEN=<team key from bootstrap>
ANTHROPIC_MODEL=tier-mid       # or tier-cheap / tier-local-small / tier-deep

Gateway access is early-access while we work with our first teams — request an invite and we'll get you the repo and a 20-minute setup call if you want one.

Managed

We run the gateway, watch the model frontier, and re-tune routing as your traffic evolves — you keep the /stats receipts. Start with the breakdown; if the numbers make the case, reply to that email and we take it from there. Zero setup on your side.

Questions we get

Does my data go through Wendl's servers?

Self-hosted: never — the gateway runs on your box, keys and spend logs stay there, local-model traffic never leaves your network. Managed: your traffic transits infrastructure we operate for you, under your own provider keys.

What happens when an agent hits its budget cap?

Requests on that key stop until the 30-day window rolls — visibly, not silently. /limits always shows where every key stands.

We already run LiteLLM.

Keep it — the plugin and the breakdown work against LiteLLM as-is. The gateway is a drop-in (same port, same key, same endpoints) whenever you want to shed the Docker/Postgres weight.

Is the routing any good?

It's backtested on your traffic before it touches production, and we only recommend a switch when it measurably wins on cost and holds quality. That's the whole point of the breakdown-first flow.

Something's broken.

Email [email protected] with what you ran and what you saw. Our setup tooling ships with a doctor that names the exact broken step — we'll usually ask you to paste its output.