You set the direction, judge the results and decide. Your coding agents run the experiments and keep the research/ folder current as they work. Research Crafter turns it into a live map, with the evidence behind each claim, that your collaborators can follow, check and steer.
Works with Claude Code, Codex and Cursor. Only the parsed research folder is uploaded, never your code.
Each one is the app as it ships today, on synthetic demo projects: real screens of each state, joined by camera moves and a drawn pointer. The terminal shots replay real CLI output. No sound.
Your agent edits research/; crafter sync uploads it as a new revision, and crafter watch keeps syncing in the background.
git status lists three files the agent changed under research/; git diff --stat research/ counts the lines.crafter sync replies: Synced revision #2 to retrieval-large, 1 status change, 1 edited.crafter watch --write syncs each saved edit: revision #3 with a status change, then #4.No new place to write research. The folder your agents already keep becomes the shared, checkable record.
As they work, Claude Code, Codex or Cursor record questions, experiments, claims, decisions, dead ends and pivots in a plain-file ARA folder in your repo.
research/ ├─ trace/ # steps, dead ends ├─ logic/ # ideas, claims │ ├─ concepts.md # LaTeX │ └─ solution/ # recipes, proofs └─ evidence/ # tables, docs
One command parses the folder on your machine and uploads it as a new revision. Unchanged folders cost nothing, so agents sync after every milestone.
$ crafter sync ✓ Synced revision #14 to curriculum-demo +1 step · 1 edited 20 steps · 7 claims · 6 concepts
Collaborators explore the map, comment on any item, ask questions with citations, and leave steer notes the agent picks up at its next task.
$ crafter inbox --write ✓ Wrote 1 steer note to .crafter/INBOX.md #3 on N18 “Two-phase mixture?” · Mai $ crafter done 3
$ curl -fsSL https://research-crafter.com/install.sh | sh $ crafter login # once per machine $ crafter init # link + first sync
crafter init also teaches your agent to sync after each milestone and read steer notes: a skill for Claude Code, a block in AGENTS.md for Codex, a rule for Cursor.
Read https://research-crafter.com/agent-setup.md and follow it.
research/ yet? crafter adopt has your own coding agent read the repo (code, notebooks, results, git history) and write the record, checked locally before anything syncs.No terminal? Upload your research/ folder from the browser.
The map, the agents, the literature, the experiments and the team: what works today, and what is being built. Anything marked Coming is not available yet.
crafter mcp; ask and go deeper in the app.crafter adopt for repos with no research record yet; or upload a research/ folder in the browser.Research Crafter does not replace the coding agent you already trust or move your work onto someone else's servers. It gives your agent a place to report to, and your lab a way to answer back.
Claude Code, Codex and Cursor keep the record as they work. Nothing to learn but a folder of Markdown and YAML.
Your laptop, a GPU box or a cluster login node: crafter sync by hand, crafter service in the background from login, or crafter install-ci for GitHub Actions on every push.
Mark a comment as a steer note. It lands in .crafter/INBOX.md; the agent acts on it and marks it done, and the note shows as delivered.
crafter mcp gives Claude Code, Codex and Cursor the map as MCP tools: read an item with its links, search, read threads and steer notes, ask, and sync, without leaving the session. crafter install-skill registers it.
The CLI parses research/ on your machine and uploads only the result. Stop syncing and you still have every word, in the open ARA format, in your repo.
14:02:11 synced revision #15: +1 step · 1 edited 14:09:40 ✉ New steer note #4 from Mai on C07: finish the third 100M seed first 14:09:40 wrote .crafter/INBOX.md 14:31:05 synced revision #16: +1 step
A result is only as good as what you can show behind it. The record keeps what was tried, what held, what failed and why, so you can defend a result rather than just report it.
Reaches the target loss 11% sooner than shuffle, ahead of linear (7%) and step (2%).
Each stage boundary causes a loss bump that takes 40M tokens to recover.
Lesson: open the pool smoothly; abrupt jumps in difficulty cost tokens.
Runs on the record itself, with or without AI. With AI on, a review also says what is supported, what is not yet, and what to check next.
Every sync is a revision you can diff, with the machine and git commit it came from.
When agents run many of the experiments, everyone needs to know what was run and why. Advisors, co-authors and new students get up to speed from the record itself, with every answer pointing back to the steps it rests on.
Why square-root pacing rather than linear?
The rule that sets how much of the data, easiest first, the learner may sample at training progress t.
Pitfall: step functions cause a loss bump at each stage boundary (N11).
A guided path of lessons and flashcards built from the project, with quizzes when AI is on.
Papers for any step, claim or concept, found on Semantic Scholar and arXiv.
Concepts and recipes carry LaTeX definitions, proofs and assumptions, rendered with KaTeX.
Without a model key, ask answers from the record itself, and explanations and diagrams are built from it.
Ablations, sweeps and failed runs as experiments, each tied to the decision it led to.
Lemmas, bounds and counterexamples as recipes and dead ends, linked to the claims they support.
Definitions and proof attempts in LaTeX, with the abandoned approaches still visible.
Derivations next to the simulations that test them, with every assumption written down.
Hypotheses and backtests with the dead ends kept, so an overfit idea is not tried twice.
Orchestra is a hosted research IDE: you describe a question, and its agents search papers, write code and run it on GPUs it provides. If you want one place to go from an idea to a result with no setup, it is built for that, and its open AI research skills library is worth installing either way. Research Crafter is for researchers who already work with a coding agent in a git repo and want their lab to follow, check and steer that work without moving it.
Orchestra details come from its public home, pricing, terms, privacy and blog pages as of 31 October 2026, linked in each row; products change, so check their site. Both products build on the idea of keeping the whole research process, dead ends included, from the ARA paper, whose authors include Orchestra's founder. Research Crafter is not a paper search engine or a metrics logger: keep your tracker for runs; Research Crafter holds what the research tried, decided and learned.
The CLI reads research/ (or ara/), parses it on your machine and uploads the result with the git commit it came from. Code, data and checkpoints never leave.
A project is visible to its members only. A public read-only link is opt-in, per project.
Research text goes to the server's AI provider only when a key is configured and a member uses an AI feature.
The research stays as Markdown and YAML in your repo, in the open ARA format. Stop syncing and nothing is lost.
Create an account, run crafter login and crafter init in your repo, and invite the people you work with.
No terminal? Upload your research/ folder instead.