Research Workbench
Menu

A look inside

The real workbench, on a made-up doctorate about supervising AI coding assistants: the papers, people and findings in the pictures are invented. Open any picture to see it full size.

Collect, read and search

Your reading in projects of your own, searched by its words and by its meaning.

The start page

The workbench start page: totals for projects, documents, references and papers, a list of projects, a note that one document could not be read, and the most recent imports.
The start page: what is in your projects, what needs your attention, and what came in last.

A folder of documents

A folder of reading notes in the workbench. Each note shows its tags, there are boxes to narrow the list by name or by tag, and the list ends with a choice of rows per page and the page number.
A folder's documents, each with its tags: click a tag to see only the documents that have it. Drag the edge of a column heading to resize the column.

Reading, with your notes beside it

A paper's working copy with its sections, two highlighted sentences, a contents list, and two notes on the highlights in a column beside the text. Above it are the paper's tags and a note on why it was kept.
Every paper gets an editable working copy. Highlight a sentence, write a note beside it, tag the paper and say why you kept it.

Searching by meaning

The search page with the question 'does seeing the tests help developers trust the assistant?'. Five passages from a paper are listed, each with its section, its page, a 'words and meaning' badge and a link to open the paper at that passage.
Ask in your own words. The passages that answer come back from every paper, each marked with how it was found and linked to its page.

Ask, with a record of every call

Answers from your own papers, and the evidence of every request to a model.

An answer from your own papers

An answer written by a model on this computer, above the search results, citing the eight passages it was shown, each a link to its paper and page, with a link to see exactly what was asked.
An answer written only from your papers, by a model on your computer, with every passage it was shown listed and linked.

The record of every AI call

The Evidence page listing two AI calls with their model, tokens, time, cost and outcome. One is opened to show the instructions the model was given, the document it was shown, and its answer.
The Evidence page: each call to a model, with exactly what it was told and shown, what it answered, and what it cost, written before the call is made.

Every call, as a tree

The Trace page for an experiment run. On the left, a tree: a search and a model call retried after the provider was overloaded, a second model writing code, a critique loop running as a sub-agent with its own two model calls, and a gate where a person approved the result. On the right, the chosen call: its instructions, the question and the answer, which cites passages by number.
Open any call to see everything it was part of: the steps, searches, retries, sub-agents and approvals, as a tree or a timeline.

Code checked before it is trusted

The Trace page in the dark theme for a code-writing run: a model writes code, a tool step compiles and tests it and reports 3 passed and 1 failed, the model revises it, and a second check passes. The first check's outputs are open on the right.
Code a model writes is compiled and tested in a sandbox, and every attempt is in the trace, failures included.

Claude, reading your vault

Claude searches, reads and cites your papers while you talk to it, on your computer or in the cloud.

What Claude read, recorded

The Evidence page listing three requests marked assistant: Claude searching the vault, Claude reading a document's first page, and Claude Code citing a reference, each costing nothing. The search is opened to show what Claude asked and the passages the vault gave, each with its paper and page, and that two matching documents were withheld by their labels.
Claude searches, reads and cites your vault while you talk to it. Every request is in Evidence, marked assistant, with exactly what the vault gave and what its labels kept back.

Connecting Claude on your computer

The Assistants tab of the desktop workbench's Settings: the entry to add to Claude Desktop's configuration, the command that adds the workbench to Claude Code, and what an assistant is and is not given.
Settings, Assistants gives Claude Desktop's entry and Claude Code's command with this computer's own path. In the cloud, Claude connects as a connector instead.

Write and plan

Papers written with their citations, and the limits and deadlines that shape them.

Writing a paper, with citations

A chapter read as a whole: three sections with in-text citations in APA 7, one with a page number, and a reference list built from them. Buttons offer the draft as Word or PDF.
Write in sections and cite from your references; the whole paper reads back in APA 7, Harvard, IEEE or Chicago, and downloads as Word, PDF, EPUB or Markdown, each with its contents list.

Every deadline in one calendar

The Calendar page in month view for October 2026, with deadlines from three projects: an ethics amendment, booking a lab, a supervision meeting at 14:00, an abstract due at 23:59, a chapter due to supervisors at 17:00, a reading group, pilot sessions and a transfer review. Buttons offer adding a deadline and opening the calendar in Outlook.
Every project's deadlines by month, week or agenda, each in its project's colour, and a file that puts them in Outlook or your phone's calendar.

A paper's limit and statistics

A paper's settings open in a drawer beside its project: a limit of 8,000 words and at least 6,000, each section's count with a box to leave it out, the citation style, and statistics including the pace of about 57 words a day to reach 6,000 by its deadline.
Set a paper's word limit, see each section's share, and the pace a day to reach it by the deadline.

The start page, in dark

The start page in the dark theme: totals for projects, documents, references and papers, three projects each with its own colour and icon, and Coming up listing six deadlines, one yesterday in red, one today, and the rest in the days ahead.
Coming up on the start page: what is due, what is late, and what belongs to which project and paper. Light or dark, as you choose.

A project's own settings

A project's settings in a drawer: its name and description, a colour chosen from named swatches, an icon chosen from a grid, and the project's deadlines below.
Each project's name, colour, icon, deadlines and sync in one drawer, without leaving the folder you were in.

In the cloud, and syncing

The same workbench in your browser, free during the beta, with your computer kept in step.

Your projects in the cloud

The cloud workbench in a browser, in the dark theme: three projects, each with its colour, icon, template and description, and an Account link in the bar at the top.
The same workbench in your browser, signed in with your account on this site: your projects, as you arranged them.

Linking your computer

The Sync settings: two privacy switches, the computer linked to a cloud account with a button to unlink it, and the project's sync mode with a Sync now button. The account and address are a test setup.
Settings, Sync: link the computer with a code, choose which projects travel, and keep everything, or only your notes, off the cloud.

The same paper in the cloud

The same paper opened in the cloud workbench in a browser, with the same tags, note, highlights and notes beside the text that were made on the computer.
Opened in the cloud workbench, the paper arrives with its tags, note, highlights and notes. Your writing is never overwritten by a sync.

Try it

Free during the beta. Register, then download it for Windows, macOS or Linux, and ask for the cloud workbench from your account. The guides say how each part works.

Something went wrong. Reload the page to continue. Reload 🗙