Local-first is easy to claim and hard to do, because the interesting question is never whether the model runs locally — it is what happens to your data on the way in and out. These five launches each answer that question differently, and all five keep the answer on your side of the wire. Two of them are about running models cheaply, and three are about the data you hand to one.
Ternary models without a build toolchain
BitRun is a zero-compile runner and terminal dashboard for 1-bit and 1.58-bit ternary language models on CPU. The research implementations usually mean cloning recursive C++ submodules and managing CMake; BitRun manages the models, runs them, and shows live wattage telemetry in a terminal cockpit. It claims 70+ tokens per second on laptops and 8B model support on Apple Silicon and x86_64.
The wattage readout is the part nobody else gives you. Ternary models exist to trade accuracy for energy, so the number that decides whether the trade is worth it is watts per token on your actual hardware, not a benchmark table. It is MIT licensed at v0.2.0-beta, and ternary models are a narrow slice of what exists — but if you want to know what 1-bit inference costs you, this is the cheapest way to find out.
A distillation proxy that tells you when to stop trusting it
Jevstiller distils a hosted model into a local one that answers like it. The part worth copying is the contract: a bound on how often the local model and the original disagree is part of the design, so you know when a local answer is unreliable instead of finding out from a customer. It runs as a drop-in proxy, so clients change only the base URL, and deployment, configuration and a Python API are documented.
That disagreement bound is the honest version of a claim every distillation tool makes. Without it you are silently serving a different model than the one you tested against. It is MIT licensed with a design document and published benchmarks, and savings depend entirely on how much of your traffic the local model actually covers — which is why the site ships a calculator rather than a headline percentage.
Your database, your key
DBJoy is a native macOS client for PostgreSQL with an AI assistant built in. You ask about your data in plain language and get answers backed by real queries, or have the SQL written for you, using your own Anthropic or OpenAI key — so the schema goes to a provider only if you send it. Around the assistant it does the ordinary client job well: schemas, safe data editing, SQL, export, ER diagrams, and schema comparison between staging and production with generated migration SQL.
Showing the query it ran is the detail that makes an assistant like this usable at all. An answer you cannot audit is a answer you cannot put in a report. It is MIT licensed, PostgreSQL only for now, with six stars — a young project, but the shape of it is right.
Parametric CAD over a local MCP server
ScratchCAD pairs a Rust solid modelling API built on the OpenCASCADE kernel with a local MCP server and an editor, so an AI assistant can build parts from Rhai scripts that produce exact B-rep solids and export to STEP, IGES, BREP, STL, OBJ and 3MF. Everything runs under Docker Compose on your machine, and nothing about the geometry leaves it.
Exact B-rep rather than a mesh is what makes this a CAD tool and not a 3D print slicer: a solid you can measure, fillet and machine from. Apache-2.0, and the README is upfront that it is under active development and the API can break between commits — clone a release tag rather than main.
Pipelines that describe themselves
Ternlux is a local-first, bring-your-own-key AI pipeline builder. You connect any source, describe the transform in natural language, and load to any destination in a visual ETL canvas. The desktop app is local-first with your own key, there is an online demo that runs the same canvas in a browser, and a community edition on GitHub. The commercial focus is the enterprise tier, which makes the free download a real starting point rather than a trial.
The visual canvas matters more than the AI here, because a pipeline you can see is a pipeline you can debug when the transform is wrong. The natural-language step removes the boilerplate, not the review.
What local-first actually costs
Five tools, five different boundaries: one keeps the model local, one keeps the key local, one keeps the schema local, one keeps the geometry local, and one keeps the data local. What you give up in each case is convenience — no managed endpoint, no hosted dashboard, sometimes no Windows build, and in every case the operational work of running it yourself. What you get is that the thing you were worried about sending somewhere never left.
