AIPOCH Open-Science: a local-first research workbench
The Electron desktop app ships 23 skills and 24 connectors, runs Python and R locally, and versions every artifact with its producing code.
In short
AIPOCH Open-Science is an Apache 2.0 desktop application that keeps AI-assisted research on the machine it runs on, combining 23 built-in skills, 24 connectors, Python and R execution, and versioned artifacts that record the code which produced them.
At a glance
- Version v0.31.1, dated September 2026; installers for macOS 12 and later, Windows x64 and Linux x64, plus a Homebrew cask.
- 23 built-in skills, 525 marketplace skills, 24 connectors and 10 specialists, per the project page.
- Licensed under Apache 2.0; roughly 4,800 stars and 289 forks on GitHub at the time of access.
- Vendor claim: 79.05 points on a comparison called BiomniBench-DA Public 50, scored with gpt-5.6-sol.
- Model-agnostic: OpenAI, Anthropic, DeepSeek and NVIDIA Build, plus Claude and Codex subscription sign-in.
AIPOCH Open-Science is an Apache 2.0 desktop application that keeps AI-assisted research on the machine it runs on, combining 23 built-in skills, 24 connectors, Python and R execution, and versioned artifacts that record the code which produced them.
A workbench, not another chat window
The pitch targets people who run experiments rather than conversations. Alongside the 23 bundled skills, the project page advertises 525 skills in a marketplace and 10 specialists. It also lists 24 connectors, with room to register custom MCP tools.
Builds ship as an Electron app for macOS 12 and later, Windows x64 and Linux x64, with a Homebrew cask on the Mac side. The current release is v0.31.1, dated September 2026.
Provenance is the actual product
Artifacts are immutable and versioned, and each one keeps a link to the code and execution history behind it. That trail is precisely what a lab notebook loses once analysis moves into a chat transcript.
A project can be exported as a portable .science package and reopened elsewhere, carrying conversation branches, selected file versions and notebook records. Imported packages open read-only.
Jobs can leave the laptop
Researchers can register SSH compute hosts and, where the cluster offers it, hand work to Slurm. The project page says the feature needs no separate activation, but it still assumes a reachable host with the right software and spare capacity.
Local-first still means traffic
Project data, settings and provenance evidence stay on disk, and API keys go into the operating system credential store where one exists. The same page concedes that model calls, web searches, connector calls and marketplace lookups all send data outward. Named providers include OpenAI, Anthropic, DeepSeek and NVIDIA Build, alongside sign-in with Claude and Codex subscriptions.
Read the benchmark line carefully
The page leads with a top score of 79.05 on a comparison called BiomniBench-DA Public 50, achieved with the gpt-5.6-sol model. That ranking comes from a self-compiled comparison in which other language models act as judges. No independent check of the figure is available.
What remains unverified
This report rests on a single source, the vendor project page on GitHub. Adoption, day-to-day skill quality and the benchmark methodology are untested here. The roughly 4,800 stars and 289 forks shown there are a snapshot, not evidence of use.
FAQ
What is AIPOCH Open-Science used for?
It is an open-source desktop workbench for research work with language models. It pairs skills, connectors and Python, R and notebook execution with versioned artifacts that retain the code and execution history behind each result.
Which models does AIPOCH Open-Science support?
The project describes itself as model-agnostic and names OpenAI, Anthropic, DeepSeek and NVIDIA Build among its built-in providers, plus custom gateways. Claude and Codex subscriptions can be used for sign-in instead of API keys.
Does AIPOCH Open-Science run on Windows and Linux?
Yes. Installers are offered for Windows x64 and Linux x64 as well as macOS 12 and later on Apple Silicon or Intel, with a Homebrew cask for the Mac. The listed release is v0.31.1, dated September 2026.
