Track new papers and releases in your field with a daily agent digest
Set up a scheduled job that pulls new arXiv papers in your categories and new releases from the GitHub repositories you follow, has a model score each item against a written description of your interests, and sends you a short digest with links. Build it in n8n, as a Claude Code routine, or with an open-source GitHub Actions project. The common problems are silent: runs that succeed but send nothing, rate limits from the arXiv API, summaries that quietly fall back to the abstract, and schedules in the wrong time zone. Add a heartbeat and check the first week by hand.
Each weekday you get one message with the few new papers and releases that match your interests, each with a link, a one-line reason it matters, and a clear note when nothing new appeared.
- Difficulty
- Intermediate
- Time
- 1-3 hours to set up
- Sources
- 6 · 6 independent
- Updated
- 2026-10-13
- Cost:
- n8n Cloud or self-hosting, or a Claude plan for routines, or free GitHub Actions minutes. Model API calls cost a little per run.
The prompt
Fill the blanks here, or leave them for your agent to ask. Pick the tab for your agent. The hand-off button gives any agent this page as markdown.
Build a daily research digest for me. Ask me before you send anything to anyone except [my email or my Slack DM]. Interests (score papers against this): [three to six sentences on what I work on, methods I care about, and what I do NOT want, for example "skip benchmark-only papers"]. Sources: - arXiv categories: [cs.CL, cs.LG]. Use the arXiv API, newest first by submittedDate, and only items from the last 24 hours. Wait at least three seconds between requests. - GitHub releases from: [owner/repo, owner/repo]. Schedule: weekdays at [08:10] in [Europe/Lisbon]. Say which time zone the scheduler actually uses. For each run: 1. Fetch new items. Skip anything already sent (keep a list of sent IDs). 2. Score each item 1-10 against my interests with a one-line reason. Keep items scoring 7 or more, at most 8. 3. For each kept paper: title, authors, link to the arXiv abstract page, and a two-sentence summary written from the abstract. If summarising fails, say "summary failed", do not paste the abstract as if it were a summary. 4. For each release: repo, version, date, and the two most important changes from the release notes. 5. Always send a message, even when nothing qualifies, with counts: fetched, scored, kept, errors. 6. If any step fails, put the error at the top of the message. Before going live, run it once by hand and show me the message. Do not post to any shared channel.
Human checkpoints: the agent stops and asks you
- Approve the interest description, sources and recipient before the schedule goes live.
- Read the first week of digests next to the raw feed to check what it missed.
- Approve any change that sends the digest to anyone other than you.
Low risk while it only sends to you. Ask for approval before it posts to a team channel, a mailing list or social media.
What you need
- A schedulerapp
n8n, Claude Code routines, or GitHub Actions.
- A model API key or planplan
For scoring and summaries.
- arXiv APIdata
Free, no key. One request every three seconds, one connection at a time.
- An email or chat account for deliveryaccount
- A short written description of your interestsdata
This is the main lever on quality. Include what to skip.
Steps
Write your interest description
Three to six sentences on topics, methods and things to skip. Add five recent papers you would have wanted, as examples.
Check: Someone else could read it and pick the same papers you would.
Choose sources and limits
Pick arXiv categories and the GitHub repos to watch. Set a cap on items per day so the digest stays short.
Check: You have a list of categories, repos, and a daily cap.
Build the workflow
Use the n8n outline, a Claude Code routine, or fork an open-source digest project. Store sent IDs so the same paper never arrives twice.
Check: A manual run fetches items and stores their IDs.
Make failures loud
Send a message every run, with counts, even when nothing qualifies. Put errors at the top. Users of one open-source digest project report runs that succeed in the log but never send email.
Check: A run with no new items still sends a message that says so.
Fix the time zone and schedule
Set the time zone explicitly and avoid weekends for arXiv. Pick a few minutes past the hour so it does not queue with other jobs.
Check: The first scheduled run arrives at the local time you expect.
Compare a week of digests against the raw feed
For five days, skim the full category listing and note good papers the digest missed or bad ones it kept. Adjust the interest text and threshold.
Check: Misses and false hits drop after your adjustments.
Ask before sharing it wider
If you want to send it to a team channel or list, run it to yourself for two weeks first, then approve the change yourself.
Check: The recipient list is only what you approved.
What goes wrong
The run shows success, but no digest arrives.
Fix: Send a heartbeat message every run and check the job log. In one digest project, a date-sorting bug meant scheduled runs never sent mail, and the run history looked normal.
The arXiv API returns HTTP 429 Too Many Requests.
Fix: Wait at least three seconds between requests, use one connection, and move the run off the busy hour. Users of a shared GitHub Actions project report 429 errors even after updates.
The summary is just the abstract copied in, because the model call failed quietly.
Fix: Make the job write "summary failed" instead of falling back to the abstract, and check your API key and model name in the logs.
The digest arrives hours late or at the wrong time.
Fix: Check which time zone the scheduler uses. One user's first run arrived about ten hours late because the schedule was in UTC.
Weekend and holiday runs are empty.
Fix: Schedule weekdays only for arXiv, as one n8n template does, or say "no new papers today" so you know it ran.
Relevance drops after an update or over time.
Fix: Re-read the interest description, add example papers, and compare against the raw feed again. Pin the version of any forked project.
A cloud routine can write to every connected tool.
Fix: Remove connectors the job does not need, and send only to yourself until you trust it.
Evidence
Two or more public accounts from different people or sites describe the same job. We summarise each source in our own words. Open the original before you rely on it.
- GitHub· Custom agentzotero-arxiv-daily, daily arXiv recommendations from your Zotero library
The project ranks new arXiv papers by similarity to your Zotero library, writes LLM summaries and emails them daily from GitHub Actions. Its README says the ranking may not match your interests and weekend runs find nothing.
TideDra · github.com - GitHub2026-07-08· Custom agentDaily run succeeds but no email ever arrives
The user says the workflow ran normally every day but never sent an email. A contributor traced it to dates in mixed formats that sorted wrongly, and submitted a fix.
bbetula · github.com - GitHub2026-05-19· Custom agentarXiv API 429
The user hit HTTP 429 from the arXiv API. Another user guessed that many forks calling arXiv from GitHub Actions at the same time caused it, and moved the run time earlier.
Archie-Zhang321 · github.com - GitHub2026-03-07· Custom agentTL;DR not summarised, abstract copied instead
The user's digests showed the full abstract instead of a summary and unknown affiliations. The maintainer pointed to a failing LLM call in the action logs.
zzezze · github.com - GitHub2026-05-21· Custom agentScheduled run arrived about ten hours late
The manual test worked, but no scheduled email came. The first scheduled run arrived about ten hours later than expected, and the maintainer explained the schedule is in UTC.
eaglezfy · github.com - Web· n8nMonitor AI research papers with Gemini-powered filtering and email summaries
The template pulls cs.AI papers from the arXiv API on weekdays only, because arXiv rarely publishes on weekends, scores them with Gemini and emails a summary of the top three.
Maxim Osipovs · n8n.io
- arXiv API user manualchecked 2026-10-13
- arXiv API terms of usechecked 2026-10-13
- n8n Schedule Trigger nodechecked 2026-10-13
- Claude Code routineschecked 2026-10-13
- ArxivDigestchecked 2026-10-13
Report how it went
Every report makes this page better for the next person and the next agent. Reports are reviewed before they count. Agents can report with the report_run tool.
Loading reports…