Claude Researcher

Claude, with a memory for your research.

If you use Claude for research, you know the frustration: every new chat is a blank slate. The paper you discussed last week, the framing you settled on, the half-formed idea you wanted to revisit — gone. Claude Researcher is a free toolkit that gives Claude a persistent record of your work, so each session picks up where the last one left off.

It runs entirely in your browser. Nothing installs on your machine, and your work lives in private accounts you own — we never see any of it.

Want to understand it first?

Copy this into any Claude chat. Claude will read the project's public source and explain it — including the caveats. Ask it anything. 

I'm a researcher considering a toolkit called claude_researcher. It says it turns Claude into a research collaborator with persistent memory. Before I decide, please read its public source and explain it to me.

Fetch these files:
  https://raw.githubusercontent.com/danparshall/claude_researcher/main/README.md
  https://raw.githubusercontent.com/danparshall/claude_researcher/main/HUMANS.md

You may inspect anything else in https://github.com/danparshall/claude_researcher — nothing is hidden.

In plain language: What does it do? What will it ask me to set up? (I understand there's a network-access setting and a GitHub connection.) Where does my data live, and who can see it? What are the real risks and downsides? Be candid — I'd rather hear the caveats now than after setup.

Ready to begin?

Copy this into a fresh Claude chat. Claude walks you through a one-time setup — about 10 minutes — and asks before doing anything sensitive. 

I'd like to set up the claude_researcher workflow (https://github.com/danparshall/claude_researcher).

Fetch https://raw.githubusercontent.com/danparshall/claude_researcher/main/README.md and find the "Quick start" section. It contains a setup prompt inside a code block. Follow that prompt exactly, as if I had pasted it here myself. It walks us through a one-time setup of about 10 minutes, with confirmation gates before anything sensitive.

You'll need a Claude Pro or Team account and a free GitHub account. Never used GitHub? That's fine — the setup assumes nothing.

What is this, really?

Claude Researcher adds persistent memory to your claude.ai research sessions: a durable record of your projects, the papers you've read, the conversations you've had, and the reasoning that came out of them. The record lives in a private repository on GitHub — think of it as a version-controlled lab notebook that both you and Claude can read and write.

It was built for researchers whose work machines are locked down — no admin rights, no installing developer tools. Everything here runs through the browser: claude.ai on one side, GitHub on the other.

What a session feels like

Suppose you work on coastal adaptation policy. You open claude.ai and start a new chat: "Let's pick up the managed-retreat work." Claude reads your profile and your project's status file, then picks up where you left off: "Last time we ended on the tension between federal buyout efficiency and state-level political feasibility. Did you find the GAO report?"

You paste in a PDF; Claude files it, extracts the text, and indexes a summary against your reading notes. You discuss for an hour. You say "good stopping point," and Claude writes a permanent record of the conversation, updates the status log, and notes what's open for next time. Next week, in a fresh chat, everything is still there.

And the record is cross-linked as it's written: your project's STATUS.md points into a running RESEARCH_LOG.md, which points to each session's conversation summary. Months later, it's easy to reconstruct why you made a decision — and easy to find the original conversation, if you ever need it.

How it works

Three pieces, all in plain sight:

Setup is a few one-time clicks, and Claude tells you exactly what to click and when:

claude.ai Settings, Capabilities pane — 'Allow network egress' turned on, with the domain allowlist set to All domains
1. Allow network access in claude.ai settings (Settings → Capabilities). This lets Claude reach the internet from inside its sandbox.
claude.ai Settings, Connectors list — GitHub Integration showing a checkmark as connected
2. Connect GitHub to Claude (Settings → Connectors → GitHub). GitHub asks you to sign in and approve.
GitHub's Create a new repository page with the visibility menu open and Private highlighted
3. Create two repositories on GitHub, setting each one to Private.
GitHub settings for the Claude app — its permissions, and Repository access with 'All repositories' and 'Only select repositories' options
4. Install the Claude GitHub App, which lets Claude save to your repositories. Choose all repositories or only select ones. The page lists exactly what the App can do.
claude.ai chat box with the permission mode menu open and Auto selected
5. Optional but recommended: set the chat's permission mode to Auto, so you aren't approving every routine step. A safety check still pauses anything that looks unsafe.

Where does my data live, and who can see it?

Your work lives in private repositories on your own GitHub account. There is no server of ours in the loop, no analytics, and no account with us. Nothing you write is sent to us.

Two companies do hold your data: GitHub stores the repositories, and Anthropic stores your chats and holds the GitHub connection you approved. Each handles it under its own terms. If your material is sensitive enough that either of those is a problem, this toolkit is not the right fit.

Where your data goes A triangle. You, in your browser, at the top. Your AI provider at the lower left, holding the chat and Claude's temporary sandbox. GitHub at the lower right, holding your private repositories. You type and upload into the chat and Claude replies; Claude's commands run in the sandbox and their output returns to the chat; the sandbox copies repositories in from GitHub and saves changes back; you can browse or download your repositories on github.com at any time. you type, upload Claude replies browse or download on github.com, anytime copies repos in saves changes back You your browser AI provider The chat your conversation, saved commands go down output comes back Claude's sandbox a temporary computer; doesn't carry over GitHub your private repositories: the permanent record
Yours

Your computer

Just a browser tab. Nothing installs. Your files stay here unless you upload them.

what you type or upload goes up; Claude's replies come back
AI provider

The chat

Your conversation, saved in your chat history. Anything that reaches the chat, Claude has seen.

Claude sends commands down; whatever they print comes back up into the chat

Claude's sandbox

A temporary computer for this chat. Claude copies your repositories here, reads files, and runs scripts. It does not carry over to your next chat.

copies your repositories in; saves changes back, through the connection you approved
GitHub

Your private repositories

The permanent record. This is what your next chat reads to pick up where you left off.

You talk to Claude; Claude works in a temporary sandbox and saves your work to your GitHub repositories, through the connection you approved. Rule of thumb: anything that reaches the chat — what you type, files you upload, and whatever Claude's commands print — Claude has seen, and it is stored with your conversation. A file sitting in the sandbox or on GitHub hasn't been seen until Claude reads it, but it is still on those companies' machines.

Will my chats be used to train AI models? That's governed by your model provider's terms of service, not by this toolkit, and those terms and their settings can change. You can ask an agent to read the current terms and translate them from lawyer to English.

Is my repository encrypted on GitHub? Yes. GitHub has stored repositories on encrypted disks since 2019. That protects against someone walking off with a hard drive. It doesn't protect against the more realistic risks: someone getting into your GitHub account, a repository accidentally switched to public, or a legal demand served on GitHub, which can read your data in order to serve it to you. Turning on two-factor sign-in for GitHub does more for you than encryption at rest.

The part that takes some trust in us is the instructions. At the start of each session, Claude downloads its working rules from our public repository, and then acts on your repositories using the GitHub access you granted. If we published bad instructions, or if our account were broken into, Claude could be told to do something you wouldn't want. We can't make that risk zero. What we can do is keep every instruction public, with its full change history, at github.com/danparshall/claude_researcher, so anyone can check what Claude is being told. The explainer prompt at the top of this page is there so Claude can look all this over for you before you decide.

Who's behind this

Claude Researcher is a collaboration between Dan Parshall (physicist-turned-data-scientist, now working on AI policy at the Canary Institute) and Andrea Lopez-Luzuriaga (economist). It grew out of our own research workflows, and we share it in the hope it's useful to yours.

The project is open source under the Apache 2.0 license. The full source, design notes, and development history are public in the GitHub repository. Found a problem? Ask Claude in any session to file an issue — or open one yourself.

Requirements