aliquot
your lab’s documents, turned into agents the lab owns.
[for labs]
an open-source agent harness for research labs. sops, codebooks and onboarding notes become agents your lab edits in a browser.
members ask in a web page and get an answer that cites the document it came from. agents can also search the literature and run code over the lab’s datasets. participant ids and anything else the lab marks restricted stay on lab hardware.
[what a lab gets]
answers from your own documents.
a steward compiles the lab’s documents into a knowledge base, word files and pdfs included, since that is where most sops live. agents search it and cite the entry they used.
research tools built in.
agents search openalex, pubmed and arxiv, list the lab’s datasets, and run r or python over them in a container with no network.
claude code or codex as the agent.
the cli gets no tools of its own and reaches only aliquot and its vendor’s api. its prompts pass the same data check as any model call. macos only for now.
written by the people who know the work.
agents and knowledge bases are forms in a web studio, with revisions, a test thread and a publish button, so nobody needs git or a deploy.
restricted data stays in the lab.
everything carries a data class. participant ids never reach a hosted model: the call is refused, and the refusal is recorded. an id inside a word file marks that source restricted too.
gaps named before publish.
a knowledge base lists the questions it has to answer. if the documents miss one, publish is refused unless a publisher signs off with a note.
a record for every answer.
every model call, tool call and refusal, hash-chained, naming the exact agent version that answered. the record opens with a summary and downloads as json, so an irb or data steward can keep a copy without running anything.
answers that keep their shape.
lists, headings and code come through as written. under each answer, a line from its record says what it read and how long it took.
runs on a laptop.
postgres and a local model through ollama, or hosted models routed by data class. no hosted service is required.
[the studio]
[the record]
[try it]
node 26, pnpm 10 and docker. no api key: the demo uses scripted models. pnpm run doctor checks node, docker, ports and the sandbox image first, and says the fix for anything missing.
git clone https://github.com/oneirum/aliquot.git
cd aliquot
pnpm install
pnpm demothe demo builds a synthetic lab and prints a link to its chat page. pnpm lab0 runs the full check: 20 questions written down before the first run, and a try at leaking a participant roster to a hosted model.
on a lab machine, one docker compose up runs the whole thing; a script backs it up and restores it.
[where it stands]
v0.5.0. tested on a synthetic lab, not used by a real one yet. it ships eleven built-in tools for search, data and code. tools a lab writes itself are not supported yet. when they come, each will run behind an operating system boundary, a separate user, a container or a vm.
next is a slack door and a router across agents. after that, studies an irb can cite.
if you run a lab and want to try it, write to us. we want to hear what breaks.
founders@oneirum.com · github.com/oneirum/aliquot · releases