The Infrastructure Layer Nobody Is Talking About
The biggest AI story this week.
The Infrastructure Layer Nobody Is Talking About
While the AI world obsesses over model capabilities, a quieter revolution is happening underneath. The real action this week is in the infrastructure layer: the tools that route, guard, and orchestrate AI requests at scale. Two projects hit GitHub trending that illustrate exactly where production AI is heading.
Bifrost (https://github.com/maximhq/bifrost) describes itself as an enterprise AI gateway, written in Go, with claims of being 50 times faster than LiteLLM. It includes an adaptive load balancer and cluster mode. Archestra (https://github.com/archestra-ai/archestra) takes a broader view, offering an enterprise AI platform with guardrails, an MCP registry, a gateway and an orchestrator, written in TypeScript.
These are not glamorous announcements. Nobody is writing breathless threads about AI gateways. But if you are building anything that touches more than a handful of users, these tools are worth understanding.
The pattern is clear. Early AI adoption meant calling an API. Production AI means building a system. Systems need routing (which model handles which request), guardrails (what does the AI refuse to do), observability (what happened and why), and orchestration (how do multiple AI calls coordinate). That is exactly what these projects are building.
LiteLLM became popular because it solved the routing problem elegantly. Bifrost is positioning itself as a faster, more capable alternative for teams hitting scale limits. Archestra is taking a more comprehensive approach, treating the MCP registry as a first-class concept rather than an afterthought.
For most readers, this is not actionable today. You are probably not running an AI gateway. But the trajectory matters. As AI moves from novelty to infrastructure, the tools that manage it become as important as the models themselves. The teams winning with AI are not just picking the right model. They are building the right system around it.
The MCP ecosystem deserves particular attention here. Both Archestra and the broader wave of MCP servers represent a bet that AI integration should be standardised, not bespoke. If that bet pays off, the way you connect AI to your tools in 2025 will look very different from the custom integrations most teams are still building today.
The Stack This Week
Bifrost Claims 50x Speed Advantage Over LiteLLM Bifrost (https://github.com/maximhq/bifrost) is an enterprise AI gateway written in Go that has gathered over 5,000 stars on GitHub. Its core pitch is performance: an adaptive load balancer and cluster mode designed for high-throughput AI workloads. The comparison to LiteLLM is deliberate. LiteLLM became the default choice for teams needing to route requests across multiple providers. Bifrost is directly challenging that position with a Go-based architecture that prioritises speed. If the claims hold at scale, it matters for any team running heavy AI traffic. The caveat is that performance benchmarks from the project itself should be treated with appropriate scepticism until independent testing confirms them.
Archestra Builds Enterprise AI Platform Around MCP Registry Archestra (https://github.com/archestra-ai/archestra) takes a broader approach than most AI gateway projects. It combines guardrails, an MCP registry, a gateway and an orchestrator into a single TypeScript platform. The MCP registry is the interesting part. Rather than treating MCP as a nice-to-have, Archestra bakes it into the core architecture. For teams standardising on MCP for tool integration, this offers a path to production that handles the surrounding concerns (security, routing, monitoring) in one place. It is early stage, but the architectural choices signal where enterprise AI tooling is heading.
Aseprite Gains MCP Integration for Game Development Workflows Game developers working with pixel art now have a new option for AI-assisted workflows. The Aseprite MCP project (https://github.com/Pand8266/aseprite-mcp-pro) connects Aseprite, the pixel art editor, to AI tools via MCP. It handles sprite, layer, frame, palette and export workflows, with explicit support for Godot integration. This is a narrow use case, but it demonstrates how MCP is spreading beyond chatbots into creative and game development pipelines. If you are building pixel art tools or working with Godot, this is worth evaluating against custom scripting approaches.
SharkMCP Brings Natural Language to Network Capture Analysis Network analysis typically requires specialised knowledge and tools like Wireshark. SharkMCP (https://github.com/Spokeswomandodgecity658/SharkMCP) connects sharkd, the daemon behind Wireshark, to LLMs through an MCP server. The pitch is that you can query network captures with natural language rather than learning packet syntax. This is a proof of concept more than a production tool at this stage, but it illustrates a broader pattern: taking specialised tools and making them accessible through AI interfaces. The network analysis space is underserved by AI tooling, making this an interesting area to watch.
I want to thank each and every one of you as we hit over 1,000 subscribers, appreciate you all and thankful for being in this awesome industry!
Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.
Gareth, founder of The Anthropic Stack (theanthropicstack.com)