Open-Source Tools From Arcnem AI

We’ve been publishing more of our tools as open source.

If you want a feel for the kinds of systems we like building at Arcnem AI, these repositories are a good place to start: image understanding pipelines, local-first similarity search, benchmark-guided coding loops, and asset analysis that stays inspectable end to end.

Arcnem Vision

arcnem-vision is the biggest picture of what we’re working on.

It starts with an image upload from a Flutter app or the dashboard, then runs a full pipeline around that image: descriptions, embeddings, segmentation, semantic search, and agent-driven processing. What makes it especially interesting is the architecture. Agent graphs are loaded from the database, MCP tools sit behind the processing layer, and the system is designed so workflows can change without hard-coding every path into the app.

If you want to see how we think about AI systems beyond a single model call, this is the repo to open first.

autoarc

autoarc is one of the most interesting repos in the set because it keeps the agent loop honest.

It is a local service for benchmark-guided code improvement. The model can propose changes, but autoarc owns the benchmark, the worktree isolation, the promotion rules, and the final decision about whether a candidate actually improves the frontier. That design matters. Instead of treating agent output as inherently valuable, it asks a narrower and more useful question: did the measured result get better?

There is also something refreshing about how small and inspectable it is. One repo, one benchmark contract, isolated experiments, durable history in SQLite.

picvec

picvec is a local-first CLI for image similarity search.

It runs ONNX vision models locally, stores embeddings in a local libsql database, and gives you a very simple workflow: initialize, embed images, and search for visually similar ones. It is the kind of tool that feels practical immediately.

texvec

texvec takes the same local-first approach for text.

It summarizes documents locally, stores both summary embeddings and chunk embeddings, and lets you search by the gist of a document or by a more specific passage. That makes it more useful than a bare document embedding alone.

omnivec

omnivec is a nice example of composition.

You upload a ZIP of documents and images, it runs local analysis with texvec and picvec, groups duplicates, builds clusters, saves structured analysis artifacts, and then writes a markdown report from those results. Most of the pipeline is deterministic and local; the LLM layer is mostly reserved for the final report. That’s a good shape for applied AI systems.

Where To Start

If any of these sound useful, check out Arcnem AI’s open-source tools on GitHub:

They’re different projects, but they share a common thread: local-first where possible, inspectable by default, and built to be genuinely useful rather than just flashy demos.