There is a persistent assumption that agent tooling means a hosted service with a usage meter. This batch of Show HN launches argues otherwise at every layer of the stack, from how model weights are stored to how you dictate an email. Some of it is ideology, but most of it is practical: the data was never supposed to leave, the latency budget does not allow a round trip, or the thing only works because it can see your actual machine.
Keeping the weights where you can reach them
At the bottom of the stack, Xet Server is a Go server that is wire-compatible with Hugging Face's Xet storage protocol, the content-defined-chunking, dedup-first layer that replaced Git LFS for large model files on the Hub. You can point the real hf CLI at it and run upload and download with byte-identical round-trips, real BLAKE3 hashing and real xorb and shard formats. For anyone mirroring weights onto their own infrastructure or working somewhere the Hub is unreachable, wire compatibility is the whole feature.
Search and dictation that never upload
Instant Search finds files by what is in them rather than by filename. You type what you remember and it returns the matching notes, PDFs, spreadsheets and documents with the relevant passage highlighted. Everything runs fully offline on Apple silicon Macs, which is the difference between a search feature and a decision about whether your documents are going into someone else's index.
Murmur is a free, open-source macOS menu bar dictation app: hold a key, speak, let go, and the words appear at the cursor in any app. It uses your own OpenAI API key rather than a subscription, which is a useful split. The app is free and MIT licensed, and the transcription cost is one you can see and control.
Isolation you can reason about
Koala runs ARM64 Linux virtual machines on Apple silicon Macs, one VM per job, service, environment or agent. Each gets its own kernel, disk and network and only the access you grant, and VMs stay up after the terminal closes. It installs through Homebrew and is at a 0.1.0 preview, but "one VM per agent with its own kernel" is a cleaner isolation story than a shared container with a permissions file.
One binary, no runtime
Caudra is a terminal coding agent shipped as a single native binary, with steerable subagents, built-in token savings, a workbench and revertible history. The parts are designed together rather than assembled, and you can open a running subagent's chat and type guidance it reads at its next turn while the main agent keeps working. Single-binary distribution sounds like a packaging detail until you have tried to install an agent toolkit across three machines.
Cutting spend without a hosted optimizer
Pith cuts output-token spend with judged replays rather than heuristics. It watches a route, replays a sample under cheaper output profiles, has a judge compare every answer against the baseline, and pins a profile only when it is as good as the baseline is with itself. When it is not, Pith says so and changes nothing. That refusal is the feature; a hosted optimizer that quietly degrades quality is worse than no optimizer.
Rooms, prototypes and canvases
Beeline is team chat for humans and agents, where a request in the room becomes an agent working in an isolated corner that produces a reviewable pull request. The published transcript shows a permissions question discussed by two people, handed to Codex, and merged with passing checks.
Kuron inverts the usual prototype flow. Instead of generating code, an LLM with tools acts as the application: you write an idea in one line, and within roughly a minute it is running on minimal screens you can use by hand. Nothing is generated per idea, so there is no codebase to maintain, and the point is to find out whether the idea is worth building at all.
Multipeer is the shared surface version of the same instinct, a multiplayer infinite canvas where your agents and your team work on ideas together and live. macOS ships first with other platforms listed next, and the site is currently taking waitlist sign-ups.
The honest trade-off
Local-first is not free. You operate the storage server, you own the VM lifecycle, you install the binary on every machine, and you are the one who notices when the index goes stale. The tools in this batch that get it right are the ones that make that cost small: one brew command, one native binary, one script tag, one menu bar app. The ones that struggle tend to be local-first in principle but still require a stack.
There is also a capability line worth being honest about. Local models and local search are excellent at retrieval and summarisation over your own material, and still weaker than frontier models at long multi-step reasoning. Mixing is normal: Murmur uses your own key against a hosted transcription model, and Enzyme builds a local index that a hosted agent then queries. The useful question is not whether a tool is local, but which part of it had to be.
The common thread is not privacy alone. It is that these tools work on things a hosted service cannot reach: your files, your hardware, your terminal, your VM.