shashishailaj/archgw

Arch is a smart gateway that sits between your AI agents and the systems they talk to. Instead of building a prototype and then spending weeks wiring up low-level plumbing, it handles the repetitive infrastructure work for you: routing prompts to the right agent, swapping between AI models, applying safety guardrails, and turning vague user requests into structured API calls. The goal is to help teams ship AI-powered apps faster by letting them focus on the product rather than the plumbing. Under the hood, it runs as a proxy server alongside your application. You can think of it as an intelligent traffic controller for AI requests. When a user sends a message, Arch inspects it, decides which AI model or internal tool should handle it, and forwards it along. You configure this behavior with a simple text file rather than writing custom code. The system supports three routing strategies: you can hardcode requests to specific models, use friendly aliases to decouple your code from specific providers, or let Arch automatically pick the best model based on plain-English descriptions of what each model is good at. A team building a customer support bot, for instance, could use this to route billing questions to one agent and technical questions to another, without maintaining that logic in their application code. The included demo shows a similar pattern: a currency exchange app where user requests like "what is the rate for GBP" are automatically matched to a live exchange-rate API, with guardrails that block off-topic or harmful inputs. A network operations team could use it to let an agent safely query device statistics or reboot hardware through structured commands. One notable design choice is that Arch is built on top of Envoy, a widely used proxy system for managing web traffic. The creators argue that prompts are just a new kind of user request, and should get the same careful handling that traditional web traffic gets. This means features like observability and security are built into the traffic layer, so you can swap models or update policies by changing a configuration file rather than redeploying your application.