Your graduate applications,
researched properly.
One workspace for every program, professor and application you are tracking. The research and the reasoning stay attached to the record, so you can read them again in November when you have forgotten why you shortlisted somebody.
- No card required
- Nothing sent on your behalf
- Runs on your own assistant
You bring the list. Your assistant does the reading.
There is no AI of ours running in the background. You connect the assistant you already pay for, and it works inside your workspace: reading pages, recording what it found and checking its own earlier work with a second pass. Your account, your model, your limits.
- 1
Make a workspace
Add the universities you already have in mind. If you have been keeping a spreadsheet, import it and keep your own ordering and columns.
- 2
Connect your assistant
One line to paste. It takes about a minute, it works the same on a laptop or a desktop, and you can revoke the access whenever you want.
- 3
Ask it to work
Say what you want in your own words. Everything it finds lands in your workspace with the page it came from, rather than in a chat you will have lost by Thursday.
$ claude mcp add --transport http gradworkbench \
https://gradworkbench.vercel.app/api/mcp --scope user
✓ Connected. 72 tools available.> research the next five professors at Michigan
From here it reads, records and checks. You watch it happen in your workspace, and every finding arrives with the page it came from.
Works with
- Claude Code
- Claude Desktop
- Claude on the web
- Codex
- Cursor
- Windsurf
- VS Code
- opencode
Applying to graduate school is a research project. Most people run it out of a spreadsheet.
Without a workspace
- A spreadsheet of 200 professors you can no longer remember researching
- Twelve open tabs of faculty pages you will read “properly” later
- No idea which of them is actually taking students this cycle
- Emails sent months ago with no record of who replied
- Deadlines living in your head, in a calendar and in a Notes app at once
With GradWorkbench
- One record per professor, and every finding carries the page it came from
- Nothing reaches a match until a second pass has checked it against that page
- Recruitment signals recorded with the date somebody actually read them
- Outreach built from a real paper and measured against your own writing
- One pipeline holding deadlines, replies, interviews and offers
One professor, start to finish.
Six steps into one record. Each step can only use what the step before it left behind, which is why the order matters and why nothing skips ahead.
- 1
Research
Your assistant works one professor at a time. The faculty page, the lab site, the publication record, the funding and deadlines for the program they sit in, and whether they are taking students this cycle.
writescandidate findings, each with its source
- 2
Check it, separately
A second pass reopens every source and rules on each finding on its own: confirmed, contradicted or not checkable at all. It is never the pass that made the claim. The platform compares the two and refuses to accept a check on your own work.
writesverified · conflicting · unverifiable
The gate
Only checked findings go any further. A candidate stays on the record, visibly a candidate, and stays out of matching, out of drafts and out of anything you would put your name to.
- 3
Match
Fit is scored against your profile using checked evidence only, with at least three distinct connection angles. Each one names the paper it rests on, so you can argue with the reasoning instead of the number.
writesa score, and the argument for it
- 4
Draft
Formal and informal outreach, built from one angle you have chosen and one real publication. Generic praise has nothing to attach itself to here, which is most of why it stops appearing.
writestwo drafts, every revision kept
- 5
Sound like you
Twelve checks for the usual tells, plus a comparison against a sample of your own writing. Anything that fails comes back naming the rule it broke.
writesa pass or a flag on each check
- 6
Send and track
You send it yourself. Replies, follow ups, interviews, deadlines and offers all hang off the same record the research is on.
writesthe thread, beside the evidence
Whatever finds it is never what confirms it.
Research produces candidates, not facts. A second pass reopens every source and rules on each finding on its own, and the platform refuses that pass if it recognises the same credential that made the claim.
- Two passes, and never the same one twice. The second is compared against whatever produced the claim, and a check on your own work is rejected rather than trusted.
- Every finding keeps the page it came from. A claim you doubt is one click from its source instead of one argument with a chatbot.
- Disagreements stay disagreements. Two sources that contradict each other produce a held finding for you to rule on, not an average and not whichever one was read last.
IllustrationTwo findings on one record
Accepting PhD students for Fall 2027.
- Department faculty pageread again
- Lab site, group newsread again
found on one pass · confirmed on another
Associate Professor on the department page. Professor on the lab site.
Held back. It stays on the record, visibly unresolved and out of every match and every draft until you settle it. A wrong title in the first line is how outreach gets deleted.
- every claim traceable to a checked finding
- no generic praise
- publication referenced is current
- name used once, title correct
- ask is specific and small
- no stacked adjectives, no manufactured transitions
- sentences average 16 words. yours average 9
- no contractions. you use about 8 per 100 words
- + 4 more, all passed
Every check is a rule, so the same draft gets the same verdict every time. You can see which one fired and why.
Sounding human is the floor. Sounding like you is the point.
A draft can clear every tell of AI writing and still read like a stranger wrote it. So the check runs both ways: a rule set for the tells, then a measurement against a sample of your own writing that you paste in once.
- Rules, not a model. The same draft gets the same verdict every time, and the verdict names which rule fired.
- It catches sentences that are all one length, the giveaway vocabulary, stacked adjectives, manufactured transitions and praise with nothing behind it.
- It measures your sentence length, how often you use contractions and how long your paragraphs run. Against your writing, not against an average. Paste a sample once.
Unattended
Queue fifteen professors. Come back to fifteen finished records.
- It holds the queue
- Work is leased and picked up where it stopped, rather than started again from the top.
- It retries once
- A transient failure goes back in the queue. A real one is marked failed and left alone.
- It stops for you
- Anything needing a human decision waits in review instead of guessing and moving on.
Everything the pipeline needs on either side of it.
Getting to a list worth researching, and getting from a reply to an offer.
Bring the sheet you already have
Import your existing list of universities and professors. Your ordering and your own columns survive the import instead of being flattened into somebody else's schema.
Build the roster before you research
Get a proposed roster for a department, keep the names worth your time and drop the rest. Research only runs on what you kept, so the expensive step is never spent on a list you never wanted.
Your own assistant, your own keys
Research runs through the assistant you already pay for, on your account. You issue the credentials, scope them to reading or writing and revoke them when you like.
Deadlines read off the source
Application deadlines, requirements and funding come from the program's own pages, and get read again rather than remembered from last cycle.
Interview prep from the same evidence
When a call lands, the preparation is built from the record you already have. Not from a fresh search the night before.
One pipeline to the offer
Stages, interviews, documents and offers sit beside the research that got you there. Nothing depends on remembering which tab you last updated.
Master's and PhD applications are not the same job.
You pick your target degree when you make a workspace, and matching, tracking and checklists weight themselves accordingly. Applying to both? Run both from the same workspace.
Master's applicants
Program first. Cost, funding, requirements, deadlines.
- Compare programs on requirements, cost and funding or assistantships
- Track deadlines and document checklists for each application
- Reach out to faculty and coordinators where it actually helps
PhD applicants
Advisor first. Overlap, recruitment signals, publications.
- Find advisors whose current work genuinely overlaps with yours
- Track who is taking students this cycle, and when that was last checked
- Build outreach that cites a real paper instead of generic praise
Three things it will not do, on purpose.
A tool that could do all of this and then send on your behalf would be a tool worth being afraid of. These are the limits, and they are not settings.
It never sends anything
No email leaves until you send it yourself. There is no bulk send and no scheduled outreach, because the moment a professor can tell an email was automated, the email has already failed.
It never applies for you
It reads, it checks, it scores and it drafts. Deciding who is worth writing to stays with you, because that decision is where the next five years of your life get chosen.
Your work stays yours
What you research and write is yours and your assistant's. It trains nothing, it reaches no other applicant, and you can export or delete the lot.
It does the reading and the remembering. You decide who is worth your time, and you send the email.
The things people ask first.
Is this for master's or PhD applicants?
Both. You set your target degree when you make your workspace, and matching, tracking and checklists adjust to it. PhD applicants get more weight on advisor fit and lab research. Master's applicants get more on requirements, funding and deadlines.
Do I need to pay for an AI subscription?
You need an assistant you can connect. Claude Code, Claude Desktop, Claude on the web, Codex, Cursor, Windsurf, VS Code and opencode all work. The research then runs on your account at your own rates, which also means your usage is visible to you and capped by you. GradWorkbench itself is free during early access.
Who actually does the research?
Your assistant does, working inside your workspace over a connection you set up and can revoke. It reads public pages: faculty directories, lab sites, program pages and publication records. It writes what it finds into your workspace and nowhere else.
How do I know the research is right?
Because nothing gets to confirm its own work. A finding like “Accepting PhD students for Fall 2027” is recorded as a candidate with the page it came from. A second pass then reopens that page and rules on it independently, and the platform refuses the check if it recognises the same credential that made the claim. Only checked findings reach a match or a draft. Conflicting ones are held for you to settle.
Does it apply to programs for me?
No. Nothing is sent on your behalf, ever. It researches, checks, scores fit and drafts. You review, edit and send everything yourself. It is a research and tracking workspace rather than an auto apply tool.
Will the emails read as AI writing?
Every draft goes through 12 rule checks for the usual tells, then gets measured against a sample of your own writing. If your sentences average 9 words and the draft averages 16, it says so. No model scores any of it, so the verdict is the same every time and it names the rule that failed rather than handing you a percentage. This page went through the same checks.
Can I bring my own list of universities?
Yes. Import an existing spreadsheet of universities and professors. It keeps your original ordering and your own columns rather than flattening them.
What does it cost?
Nothing during early access, and no card is needed to start. Paid plans come later, once the workflow has been proven on real applications.
Start with one professor.
Make a workspace, connect the assistant you already use and see what one properly checked record looks like. Free during early access. No card.