Evaluating LiteLLM alternatives

What to ask any gateway — including us

1. Does governance match your org chart?

   Ask Do budgets stack across organization, team, user, model and key? Do they fail closed? Can one team exhaust another team's capacity?

   Why it matters Per-key limits work for one team. Once spend has to be attributed to cost centres and capped per team, a key is the wrong unit.

   Wäg Budgets stack from organization through team, user and application to key, with fair-share caps, and a breach rejects the request before the provider is called.

2. Is policy enforced in the request path?

   Ask Does a rule apply to an application that never integrated anything? Or only where somebody remembered to add a hook?

   Why it matters A rule that applies only where it was opted into is advice, not a control — and the application nobody registered is the one that leaks.

   Wäg An ordered policy chain — pre-call, mid-stream and post-call — runs on every request that reaches the endpoint.

3. What is the runtime footprint?

   Ask What ships in the image? An interpreter, a package manager, a dependency tree resolved at deploy time?

   Why it matters Everything in that image sits in front of every AI request. It is supply-chain surface and operational surface, and it grows with the gateway's importance.

   Wäg One distroless, non-root Rust binary. The benchmarks page has the image size, memory and startup figures, measured side by side. See the benchmarks

4. Are routing decisions explainable?

   Ask Does the same request always produce the same decision? Does the response say which model served it, and why?

   Why it matters If nobody can explain why a request went to the expensive model, nobody can fix it — or defend it to finance.

   Wäg Intelligent routing scores requests structurally, so the same request gets the same analysis. Every response names its model in x-waeg-routed-model; intelligent routing adds x-waeg-route-reason.

5. Does it govern agent tool calls?

   Ask How are MCP servers authenticated? Can tools be allow-listed per agent? Is every hop of an agent run budgeted and recorded? Can a call be held for a human?

   Why it matters One user turn becomes dozens of model and tool calls, some with real credentials. A gateway that sees only the model calls sees the harmless half.

   Wäg Per-server auth, tool allow-lists and pinning, run limits, approvals by a named person — in Slack or Teams too, with Enterprise — and per-hop budget and audit.

6. What does it produce for audit?

   Ask What would you hand a compliance team six months from now — traffic logs, or a record of what was decided and why?

   Why it matters Since 2 August 2026, EU AI Act transparency obligations and GPAI enforcement carry penalty exposure. Logging and traceability have to be demonstrable.

   Wäg A decision record per request — model, reason, policy verdicts, cost, actor — in an open-source, signed evidence ledger you check offline with waeg ledger verify. Enterprise adds evidence packs mapped to the AI Act and DORA, with gaps flagged, and AI content marking.

7. Who operates the control plane?

   Ask Does a self-hosted data plane report to a vendor-operated control plane? What leaves your perimeter? What happens if the vendor is acquired?

   Why it matters A vendor above the request path is a dependency in front of production, and its continuity becomes yours.

   Wäg You do. There is no vendor control plane, no telemetry and no phone-home — both planes run in your infrastructure.

8. What is the vendor's jurisdiction and licence?

   Ask Where is the vendor incorporated? Does the software phone home or enforce a licence at runtime? Can the licence change under you?

   Why it matters The licence decides what you can keep running if the relationship ends. The jurisdiction decides who else can make demands of your supplier.

   Wäg Bitkaio Ltd, UK. The gateway's production path is Apache-2.0 — routing, HA, caching, failover, budgets, policy and MCP. The licence is a signed offline token with no licence server; Enterprise adds identity, audit evidence and agent payments.

Where LiteLLM is ahead

A guide that only lists our strengths would not be worth reading. LiteLLM has:

• A longer list of providers, including long-tail ones Wäg does not cover yet.
• In-process Python callbacks, if your team extends the gateway in Python.
• A large community and years of production use. Wäg has not had its public release yet.
• A SOC 2 Type II report and IP allowlists in its enterprise tier. Wäg has neither today.

Three architectures, not twenty products

Products in this category change every quarter. The architecture underneath them does not, and it decides most of the questions above.

Same API, one layer up

Wäg is a self-hosted AI control plane with the gateway included. It speaks the same OpenAI-compatible API and imports LiteLLM configuration, so trying it costs one base URL — and you can run both side by side until the numbers agree.