The Layer Under the Agent: Gateways, Retrieval and Local Runtimes
Most of the attention goes to the model and the chat surface, but the launches that quietly decide whether an agent works in production sit underneath: how a call reaches a model, how the agent finds the right context, and how much of it runs on hardware you control. The recent batch of Show HN projects is heavily weighted toward exactly that plumbing, and several of them are more interesting than the assistants built on top.
Here are nine of them, grouped by the job they do.
One door for every model
Agents that start as a demo with a single provider tend to grow into a mess of SDKs, keys and per-provider error handling. Experiential is an open source gateway and router that puts hosted, bring-your-own-key and local models behind one OpenAI-compatible API. Convenience is the least of it. Once real traffic arrives, teams want two other things: control over which users and agents may call which models and how much each can spend, and a way to turn that traffic into something better. Experiential aims at both, letting production logs become a custom router or a specialized model you own rather than rent.
The direction of travel matters here. A gateway that only proxies is a commodity, and commodities get replaced the moment a provider ships the same feature. A gateway that learns from what your agents actually ask, and routes on that, becomes infrastructure you would think twice about removing.
Retrieval that reads instead of matches
Chunk, embed, rank by cosine similarity, hope. The standard vector pipeline has well-known failure modes on long, structured documents, where the right answer depends on where a passage sits rather than on the words it happens to use. PageIndex takes another route: it builds a tree index over the document and lets a model reason over that tree, so retrieval becomes context-aware and behaves more like a person skimming a report than a nearest-neighbour lookup. It ships a Python SDK with local and cloud modes, and a file-level index layer that lets it reason across an entire corpus instead of one file at a time.
The trade-off is stated honestly. Reasoning costs more tokens per query than a vector lookup, and quality tracks the model doing the reasoning. What you buy in exchange is a retrieval path you can actually read when it fails, instead of a similarity score that gives you nothing to debug.
Models on hardware you own
Two projects push inference back onto the machine in front of you. Nobodywho runs an open-source inference engine on devices and servers, with SDKs for Rust, Python, Kotlin, Swift, Flutter, React Native and Godot. The promise is no per-token costs, no data leaving the machine, and no dependency on a network connection, which for regulated teams is the only version of AI they are allowed to deploy.
Ollaya is narrower and, for that reason, easier to adopt this week. It downloads and serves small open decision models locally and answers typed questions: you define the options, the model scores each one, and you get the top answer with calibrated scores in milliseconds. If your agent currently asks a general model to classify a support email and then parses the prose back out, this replaces the entire ritual with a structured call.
Programs instead of prompt strings
Prompt strings rot, and nobody notices until the model changes underneath them. Imp ports the DSPy programming model to the BEAM: you declare what each language-model step takes and returns, choose how it should think, and let an optimizer tune the program against examples of good output. You get signatures, modules, optimizers, agent loops and retrieval, all running under OTP supervision. For Elixir and Erlang teams it removes the last good excuse for keeping the model glue in another language.
A web that behaves the same twice
The browser is still where agents are most fragile, mostly because every action is re-derived from scratch on every run, so results vary and context fills up. Reduck attacks this by making the agent write and reuse scripts instead of clicking. A script returns the same structured output every time against a clear contract, runs in your own Chrome with your existing logged-in sessions, and burns no tokens while it executes. The expensive step moves from per-action reasoning to a one-off authoring step, which is where you want it.
Wiring agents to systems that already exist
Three more projects do the unglamorous work of connecting agents to tools people actually use. OpenFlow is a visual builder for multi-agent workflows: drag role cards onto a canvas, wire a planner to an architect to a coder, save the pipeline and run it with real parallel agents on a forked opencode engine. Cloudflare cf CLI rebuilds the Cloudflare command line for agents, with commands an agent can discover through built-in search, JSON output condensed to save context, and a single configuration file for the platform. Konnect plays the same role for hardware: a Rust binary that exposes hundreds of KiCad schematic, layout, routing and design-review operations to any model over MCP, loaded as on-demand toolsets so the context stays manageable.
The pattern
Not one of these nine is a chat interface. They are the parts that decide whether an agent is cheap or expensive, correct or confidently wrong, portable or locked in. When the models themselves are close enough that swapping one is a config change, the leverage moves to the layer underneath, and that is where this batch of launches is concentrated.
