Getting Started

The Research Workbench: How the Tools Connect

One drawing of how the host, its connections, the models and your own records fit together, with what each part does and where to set it up.

See where each tool sits before you install any of them.

The research workbench is the set of tools that lets an AI work on your own research material while you stay in control of it. Your library, notes and papers stay in files you own. The AI reaches them only through the connections you add, and you decide what each connection can see.

This page draws the configuration the systematic-review module ran. Cherry Studio is the host because most participants arrive from a chatbot, and it lets them compare models side by side before an agent touches their files (Tools & budget). Claude Code or another agent host can take the same position, with the same connections beneath it.

+---------------------------------------------+
| Researcher                                  |
| question / judgment / checks                |
+----------------------+----------------------+
                       |
                       v
+----------------------+----------------------+
| Host: Cherry Studio                         |
| chat / agents / knowledge base              |
+---------+-----------------------+-----------+
          ^                       ^
          |                       |
          v                       v
+---------+----------+  +---------+----------+
| Tools through MCP  |  | Models through API |
| Zotero MCP         |  | gateways:          |
| filesystem MCP     |  |   OpenRouter       |
|   (vault only)     |  |   Haimaker         |
| Research Memex     |  | local model        |
|   MCP server       |  |   (optional)       |
+---------+----------+  +--------------------+
          ^
          |
          v
+---------+----------+  +--------------------+
| Durable record     |  | Discovery          |
| Zotero library     |<-+ Research Rabbit    |
| Obsidian vault     |  | OpenAlex           |
| papers as Markdown |  +--------------------+
+--------------------+

Read it from the top. You direct the host; the host calls the tools and the models, and each call returns a result, so those arrows run both ways. The models never open your files. They see what the tools hand the host, and what the model writes back reaches your records only through a tool. Discovery tools feed the record from outside, as papers you decide to keep.

What each part does

Component	What it does	Why it matters	Guide
Researcher	Brings the question, makes the judgments, checks the output	The only part that knows what the review is for	Session 1
Cherry Studio	Holds the conversation and the agents, and calls the tools and models below it	One window to compare models before an agent touches your files	Cherry Studio
Zotero MCP	Lets the model search and read your Zotero library	Citations come from your library, not from the model's memory	Zotero MCP
Filesystem MCP	Lets the model read and write files in the folders you allow	Scope it to the vault folder only, so the rest of your disk stays out of reach	MCP Explorer
Research Memex MCP server	Serves this site's pages and its three roles to the model, at `research-memex.org/mcp`	The model can look up a method or run a role without you pasting pages in; the roles run on your own model	How AI reads this site
OpenRouter, Haimaker	Gateways: one account reaches models from several providers. The course buys credits once on OpenRouter and issues each participant a named key with its own spend cap. Haimaker describes itself as an AI API gateway to many providers' models through one OpenAI-compatible endpoint; the second run used it too	Everyone can run the same task through different frontier models without opening accounts	API Keys
Local model	Runs a model on your own machine	Optional, free and private, and weaker than frontier models	Free and low-cost options
Zotero library	Holds your sources and their bibliographic records	The citation record that outlasts any conversation	Zotero
Obsidian vault	Holds your notes, protocol and the AI's output as Markdown files	The place decisions are written down and can be found again	Obsidian
Papers as Markdown	Converted text of each PDF	A model reads and quotes Markdown more reliably than a PDF	PDF to Markdown
Research Rabbit	Finds related papers through citation networks	Widens the search from the papers you already trust	Research Rabbit
OpenAlex	An open index of the literature with an API and a CLI	A search an agent can run and you can rerun	OpenAlex

A round trip

You ask. Paste your protocol into the host with the question, so the model works to your criteria rather than its own.
The model reads. It searches your library through Zotero MCP and opens the Markdown papers through the filesystem MCP.
It writes. The output goes into your vault as a note, with a locator, a page or a section, for every claim it takes from a paper.
You check. Hold each locator against the paper. If the output reads generic, run the Output Mirror on it. Then record the decision in the vault, including the ones where you overrule the model.

Minimum, recommended, full

Minimum: Zotero, Cherry Studio and one model provider. Enough to ask about papers you paste in and to compare models.
Recommended: add the filesystem MCP, Zotero MCP, Obsidian and PDF-to-Markdown conversion, so the model reads your library and papers and writes into your vault.
Full: add a local model, a second gateway, the Research Memex server and agent teams, as Session 4 runs them.

For a guided setup, Carrel configures the machine with you; for the order to install things in, follow Quick Start.

cite this page

Lin, X. (2026). The Research Workbench: How the Tools Connect. Research Memex. https://research-memex.org/docs/implementation/research-workbench

@misc{docs-implementation-research-workbench-2026,
  author = {Xule Lin},
  title = {The Research Workbench: How the Tools Connect},
  year = {2026},
  howpublished = {\url{https://research-memex.org/docs/implementation/research-workbench}},
  note = {ORCID: 0000-0001-7885-4194}
}

one renderingthe source remains