An AI gateway is the one path every model, tool and agent call takes out of your organization. Skyfix runs that path as a single pipeline: identity, policy, inline AI Firewall, budget, routing and evidence, in the same order for a chat completion, an MCP tool call or an A2A task.
The trace of one request shows the task Garuda recognised, the rule it matched, every route it ranked with its price, and why the top one was ruled out.
Garuda read the latest message and recognised code generation with 99% confidence, above the rule's 85% threshold.
The matched rule's profile ranked every allowed route on tier and tags and on price; the decision note says the better-suited routes had no allowed model.
reasoning-xl, the only Frontier route, ranked first but was ruled out: this caller's API key is not allowed to use it.
With Frontier ruled out by the key, the next route was a Structured output model. Two providers offer assistant-large; the cheaper one served the request, with no fallback.
One control plane
Hard gates run before any preference. The router only chooses among what policy already allows, and the same six stages apply to every protocol.
Identity
Policy
AI Firewall
Budget
Route
Evidence
Most organizations end up with one proxy for model traffic, a second for tool servers and nothing at all in front of agents that call other agents. Each one sees a different slice of the same request, with its own keys, its own logs and its own idea of policy. Skyfix replaces those with one gateway that treats a chat completion, an MCP tool call, an A2A task, a skill step and a batch line as the same kind of thing: a governed request.
Every request is authenticated against a scoped key or workload identity, checked against a signed policy bundle the gateway verified locally, inspected inline by the AI Firewall, priced against its budget, routed, and recorded in the decision ledger. Hard gates such as regions, provider allow and deny lists, budgets and guardrails run before any preference, and a request that fails a gate never reaches a provider.
Applications keep the client they already have and change the base URL. The gateway serves an industry-standard chat-completions API and other widely used model API formats, plus embeddings, rerank, moderation, image, audio and realtime endpoints, with server-sent events for streaming. Model aliases are the contract: an application asks for a capability name, and platform teams decide which provider and model back it.
Batches run through the same path. Each line of a batch is inspected, routed and priced like a live request, groups of lines are recorded as sent before dispatch so a restart never sends a line twice, and a batch is visible only to the key that created it. A prompt registry lets the gateway render a named, versioned prompt before inspection, so a template change is reviewed once and the AI Firewall still sees the rendered request.
Public model providers, bring-your-own-key accounts, cloud-hosted model services and private or self-hosted endpoints in your network are configured once and addressed through aliases. Each provider is connected over HTTPS with a connection test, health checks and a circuit breaker; a load-balancing strategy such as fallback order decides what happens when the preferred provider is unavailable.
Failover never escapes policy. A fallback provider must satisfy the same region, data-residency and allowlist rules as the primary, and the ledger records which providers were skipped and why. A private model in your own cluster is a first-class provider: it gets the same alias, the same inspection and the same evidence as any public endpoint.
Tools and agents are where most AI risk now lives, and they get the same six stages. The MCP Gateway keeps a registry of tool servers whose catalogs are reviewed and approved, applies per-tool approvals and argument policies, limits calls, and authorizes clients with OAuth 2.1. The A2A Gateway publishes a signed Agent Card for your inbound agent, accepts tasks from allowlisted partners within their scope and rate limit, and federates outbound tasks to external agents.
A tool call or an agent task passes the AI Firewall on its arguments and results, draws from the same budgets as model calls and lands in the same hash-chained ledger. Both implementations are tested against the official conformance suites.
When an application asks for 'auto', Garuda, the task router, classifies the request and ranks the models that are still eligible after every gate has run. It can prefer a lower-cost model for a classification task and a stronger one for long-form reasoning, but it cannot pick a model that policy denies, a region that is not allowed or a provider whose budget is spent. Router heads are signed and verified like any other artifact; when none is present the gateway falls back to baseline routing. Every decision records the route reason, the models skipped and the fallback reason.
Rate limits apply per key and per team on requests, tokens and concurrency, enforced or measure-only, with standard rate-limit headers in every response. Semantic and response caches are administered per installation, so a cached answer is served only where policy allows and is recorded as a cache hit in the ledger. Batches share the same limits and drain behavior: when a gateway is draining or a limit is hit, a line goes back to pending until the retry window passes.
07
Versioned, approved Agent Skills with progressive disclosure let teams package what an agent may do; every skill step runs as a fully governed request. Durable workflows add checkpoints, retries, compensation, agent memory, human approval steps, environment promotion with dependency checks and signed webhook triggers, so a long-running agent process can be paused, resumed and audited like any other change.
FAQ
An AI gateway is a single, governed path between your applications and the models, tools and agents they call. It authenticates the caller, applies policy and inspection, enforces budgets, routes the request and records what happened. Skyfix runs that path as one pipeline for LLM, MCP and A2A traffic.
An LLM gateway only sees model calls. An AI gateway also governs the tool calls agents make over MCP and the tasks agents exchange over A2A, with the same identity, policy, firewall, budget and evidence stages. In Skyfix the LLM gateway is one part of the AI gateway.
Yes. The gateway serves an industry-standard chat-completions API and other widely used model API formats behind one base URL, so most applications change the endpoint and the model name to an alias and keep the rest.
Yes. Private endpoints in your network and self-hosted models are connected like any provider, with a connection test, health checks, a circuit breaker and an alias, and they pass the same policy, inspection and evidence as public providers.
A load-balancing strategy such as fallback order selects the next eligible provider. The fallback must satisfy the same policy as the primary, the circuit breaker on the failed provider opens until a probe succeeds, and the ledger records the providers skipped and the reason.
Next step
Deployed in your Kubernetes, under your keys.