All features
Intelligent Routing

Route to the cheapest model that can actually do the job

Save money without breaking tools. On the intelligent strategy, Wäg scores each request structurally, applies hard capability floors, then picks the cheapest model that clears them — and can tell you why in the response headers. If a provider is restricted or goes down, routing fails over to the next, so you keep serving instead of going dark.

"On the intelligent strategy, ambiguous short prompts upgrade a tier — they never silently downgrade quality."

Structural scoring

Each request is scored on tools, vision, context length, reasoning cues, and multi-turn depth — not a guess from a black-box classifier.

Hard capability floors

Requirements become floors: a vision request can never land on a text-only model. On the intelligent strategy, ambiguous short prompts upgrade a tier rather than downgrade.

Cheapest eligible wins

Among models that clear the floor, Wäg picks the cheapest. You save money on the easy requests without breaking the hard ones.

Explainable decisions

Responses can carry x-waeg-routed-model and x-waeg-route-reason. The scoring is deterministic: the same request always gets the same analysis.

Stronger-first fallbacks

On the intelligent strategy, fallbacks are ordered stronger-first, so a retry after a failure never quietly degrades the answer.

Six strategies

weighted, latency, cost, round_robin, intelligent and model-driven — set for the platform or per organization. An optional routing model can pick the tier, never below hard requirements.

Language-aware

The router detects the request language, nudges non-English requests slightly toward quality and prefers deployments tagged for that language. Untagged ones stay eligible.

Session affinity

Coding agents and chat apps stay on the deployment that served their last turn, so the provider's prompt cache keeps paying off. A pin yields when a turn needs more.

Learns, if you let it

Optional Thompson sampling learns from successes and failures among eligible deployments. A new key starts from what its organization has learned; floors still apply.

Intelligent Routing
Fallback chains, configured per model alias in the console.

Step by step

  1. 01

    Analyze

    The request is decomposed structurally — modality, tools, vision, context size, and reasoning signals become an explicit capability profile.

  2. 02

    Floor

    Required capabilities are turned into hard floors. Any deployment that cannot meet them is excluded before cost is even considered.

  3. 03

    Select

    Among eligible deployments, the strategy picks one: the cheapest, the fastest, by weight or in turn — set for the platform or for each organization.

  4. 04

    Explain & fall back

    The chosen model and the reason are recorded and can be returned as headers. On failure, ordered fallbacks retry — stronger-first on the intelligent strategy — with circuit breakers in play.

response headers
# Wäg can tell you what it did (illustrative values)
x-waeg-routed-model: openai/gpt-4o-mini
x-waeg-route-reason: score 18 · conf 0.91 · min_tier economy · no tools, short ctx
x-waeg-cost-usd: 0.00041
Part of the control plane

Intelligent Routing is one of six control surfaces enforced in the same request path, applying to every application behind the endpoint — including the ones that were never instrumented.

See how they combine

One endpoint. Every model. Complete control.

Start with the open source

Run it in your own infrastructure

Everything you need to run a control plane in production is Apache-2.0 and stays that way — routing, HA, caching, failover, budgets, policy, MCP, jobs and observability. Self-hosted, no account, no phone-home.

$ docker run -p 4000:4000 <registry>/waeg-gateway:<version>

Not a working pull: the images are not public yet. Design partners receive the registry details privately.

Deploy in 5 minutes
Design partnership

Shape it as a design partner

Wäg is not publicly released yet. We are working with a small cohort of teams running multi-provider AI in production, and serious about governance and sovereignty, to harden the control plane on real workloads.

  • Every Enterprise capability included, at no licence cost
  • Hands-on migration — we do it with you, including from LiteLLM
  • A direct line to the engineers building it, and real influence on what ships next
Apply to the program