conference_agent

https://conferenceagent.vercel.app

An AI agent that automatically compiles a table of major conferences. Includes website links, color-coded prior and upcoming submission deadlines and dates, and biweekly updates (every two weeks). Keeps it searchable in a web table, and exports each conference's deadlines and dates as a subscribable calendar feed (.ics). Discovery is seeded across medicine, genomics/bioinformatics, and data science; see TAXONOMY.md for the field map and SEED_CONFERENCES in config.py to add more.

Install

conda activate conference_agent
pip install -e ".[dev]"        # core + test tooling
pip install -e ".[discover]"   # discovery agent helpers
pip install -e ".[web]"        # FastAPI web table + calendar feed

Usage

conference-agent discover --subcategory radiology --email   # find + store (+ email summary)
conference-agent list                                    # print the stored table
conference-agent serve                                   # launch the web table at :8000

Add conferences manually

To enter or correct dates by hand — no API, no discovery agent — use conference-agent add. The flags mirror the table columns (the only extra is --url, the link behind the conference name); dates are ISO YYYY-MM-DD, and the submission/conference month columns are derived from the dates automatically. By default only the fields you supply are written, so an existing series keeps the rest of its data:

# Add (or update) a single conference via flags.
conference-agent add \
  --conference "RSNA - Radiological Society of North America Annual Meeting" \
  --subcategory radiology "machine learning" \
  --format abstract poster oral \
  --location "Chicago, IL" \
  --attendance 39000 --attendance-year 2025 \
  --remote-option hybrid \
  --cost "\$1,095 (member, early-bird)" \
  --abstract-due 2026-05-06 \
  --conference-dates 2026-11-29 2026-12-03 \
  --url https://www.rsna.org/annual-meeting

The Size column (large / medium / small) is derived automatically from --attendance — there is no size flag to set. A figure of 1,000+ is large, 100–999 medium, under 100 small; with no attendance, size is left blank. (The thresholds live in conference_agent/models.py.)

--conference takes the table's first column verbatim — ACRONYM - Full Name. A bare acronym (or the ACRONYM - Full Name form) that matches a row already in the table counts as a match: the command shows that entry and asks you to confirm before updating it, so a typo can't silently overwrite an existing series. Pass -y/--yes to skip the prompt (required when running unattended). If the series already exists, the command updates only the fields you pass — add --overwrite to replace the whole row instead, clearing anything you omit.

To load several at once, point --csv at a file whose header columns are the same column names as the flags — conference, subcategory, format, location, attendance, attendance_year, attendance_source, remote_option, cost, abstract_due, paper_due, conference_dates, url — with one conference per row (the derived size and category columns, as in a table export, are accepted but ignored). conference_dates is a single cell holding the start and (optional) end date separated by a space, and a multi-value subcategory or format (any of abstract / paper / poster / oral) is quoted so its comma stays inside one cell. The broad category is derived from the subcategory automatically, so you never set it. The raw stored field names (acronym, name, upcoming_start_date, …) are also accepted, so the web table's "Export CSV" re-imports unchanged.

conference-agent add --csv my_conferences.csv

conference,subcategory,location,attendance,remote_option,abstract_due,conference_dates,url
RSNA - Radiological Society of North America,radiology,"Chicago, IL",45000,hybrid,2026-05-06,2026-11-29 2026-12-03,https://www.rsna.org/annual-meeting
SPR - Society for Pediatric Radiology,"radiology, pediatrics","Austin, TX",1200,in-person,2025-12-01,2026-05-12 2026-05-16,https://www.pedrad.org
MICCAI - Medical Image Computing & Computer Assisted Intervention,"radiology, machine learning","Daejeon, South Korea",3500,hybrid,2026-03-05,2026-09-23 2026-09-27,https://miccai.org

Scheduled refresh

Discovery re-runs on three cadences: daily the targeted auto-check (--cadence due — only series whose next edition is plausibly about to be announced, each re-checked at most every RECHECK_INTERVAL_DAYS (14) days), weekly the flagship fields (--cadence weekly), and monthly the rest (--cadence monthly). The "due" auto-check is gated by a last_checked column, so it needs a persistent CONFERENCE_DATABASE_URL to work across runs.

These cadences are scheduled in the AWS SAM stack via EventBridge Scheduler rules that invoke a refresh Lambda (web/refresh_handler.py); enable them by deploying with EnableScheduledRefresh=true (see DEPLOY_AWS.md). They replaced the former .github/workflows/{weekly_update,monthly_update,auto_check}.yml cron workflows. You can also run any cadence locally with daily_update.py (below).

See TAXONOMY.md for the field map and cadence policy.

Manual refresh (running outside the auto-check window)

If you want to refresh on demand (outside of the scheduled refresh) — run daily_update.py yourself. By default these use the claude-code backend (the local claude CLI / your subscription), so no ANTHROPIC_API_KEY is required:

# Refresh a single field immediately, ignoring the staleness window entirely.
# --subcategory overrides the cadence selection, so the 6–12-month window and the
# 14-day re-check interval do NOT apply — it re-discovers that field right now.
python scripts/daily_update.py --subcategory genomics

# Refresh several fields at once.
python scripts/daily_update.py --subcategory genomics --subcategory radiology

# Refresh everything (every seeded field), also bypassing the due-window gating.
python scripts/daily_update.py --cadence all

# Run the same auto-check the scheduled job runs (only series currently "due").
python scripts/daily_update.py --cadence due

# Add --no-email to skip the summary email; add --backend api to use the
# metered Anthropic API instead of the local claude CLI.

For a one-off discovery of a single field without the refresh wrapper:

conference-agent discover --subcategory genomics --email

REST API (for agents and scripts)

The web layer is a read-only, credential-free REST API, so any AI agent or script that can make an HTTP GET can pull the data directly — no MCP server or custom integration required. Interactive docs are at /docs.

Endpoint	Purpose
GET /api/search?q=<query>&format=json	Boolean search; returns the conference rows as JSON (format=csv for a CSV export).
GET /api/fields	Self-describing list of queryable fields, aliases, and controlled vocabularies — call this first to learn the query grammar at runtime.
GET /api/calendar.ics?q=<query>	The selected conferences as a subscribable iCalendar feed.

The q parameter uses the same boolean query language as the web table (see Boolean search below); an empty q matches everything. An agent that needs to construct valid queries should read /api/fields first, since it returns the exact field names, aliases, and allowed values the parser accepts.

Do I need an MCP server?

Generally no. The REST API above is public, self-documenting (via /api/fields), and already structured for programmatic use, so scripts, the Claude Agent SDK, web-fetch-capable models, and the discovery pipeline can all consume it as-is. An MCP server would mostly be a thin wrapper over these two endpoints; it is worth adding only to give MCP-client users (e.g. Claude Desktop / Claude Code) a one-click, typed search_conferences tool without each writing their own HTTP integration. For everything else, point the agent at /api/fields and /api/search.

How it works

1. Discover — discover.py runs an Anthropic web-search loop (research) then a structured-output call (extraction) to produce typed Conference records. A seed list of well-known conferences (SEED_CONFERENCES in config.py) bootstraps the search; the reference pages those seeds were compiled from are recorded alongside it in SEED_CONFERENCE_SOURCES.
2. Store — records are upserted into a SQL database, idempotent on the acronym, so re-running discovery rolls a newly announced edition into the "upcoming" columns instead of duplicating the row.
3. Search — a web table (conference-agent serve) supports a boolean query language, a "Subscribe (.ics)" calendar feed, and a per-row "📅 cal" button that downloads that conference as a calendar file.
4. Calendar — each conference's upcoming abstract deadline, paper deadline, and conference dates are served as a credential-free iCalendar feed (GET /api/calendar.ics) that mirrors the active search. A user subscribes from any calendar app (Google "Add by URL", Apple/Outlook "Add from URL") or downloads a one-off .ics — no sign-in, no Google account. Each event carries a stable id, so re-fetching updates events in place rather than duplicating, plus reminders four weeks, one week, and one day ahead.
5. Notify — an optional email summarizes a discovery / daily refresh.

Boolean search

The web table accepts queries like:

• (virtual OR hybrid) AND size:large
• subcategory:radiology NOT cost:*
• category:medicine AND size:large
• format:poster AND format:oral
• upcoming:>=2026-06-01

Fields support field:value, field:"quoted value", presence tests (field:*), date comparisons (>, >=, <, <=, =), AND/OR/NOT, and parentheses.

Plain-English search (optional, local LLM)

The "✨ Ask" box turns a plain-English request — "big radiology conferences between September and January" — into the boolean query above, then drops it into the search box (visible and editable) and runs it. The translation runs on a free, local Ollama model — no API key and no external network call — and the model's output is validated against the real parser (with one repair round) before it is shown.

It is entirely optional: if no local model is running, the box reports that and the manual boolean search keeps working. To enable it:

# one-time: install Ollama, then pull a small instruction model
ollama pull qwen2.5:1.5b     # the default; tiny and fast
ollama serve                 # if not already running as a service

GET /api/translate?q=<plain English> returns the compiled {"query": ...}.

The default qwen2.5:1.5b is the lightest option and handles direct requests well, but it can stumble on multi-step phrasing — e.g. a wrap-around month range like "September through January of any year." A larger model translates those more reliably; point CONFERENCE_NL_QUERY_MODEL at one you have pulled:

CONFERENCE_NL_QUERY_MODEL=llama3.2:3b conference-agent serve   # or qwen2.5:7b

Configuration

• ANTHROPIC_API_KEY — required only for the api discovery backend. The default claude-code backend uses the local claude CLI / your Claude Code subscription and needs no key (for unattended runs it uses CLAUDE_CODE_OAUTH_TOKEN from claude setup-token instead).
• SMTP_HOST / SMTP_PORT / SMTP_USER / SMTP_PASSWORD and CONFERENCE_NOTIFY_EMAIL — optional, enable the summary email.
• CONFERENCE_DATABASE_URL — optional, overrides the default SQLite location.
• OLLAMA_BASE_URL (default http://localhost:11434), CONFERENCE_NL_QUERY_MODEL (default qwen2.5:1.5b), and CONFERENCE_NL_QUERY_TIMEOUT — optional, configure the local model used for plain-English search.

Deploy (static site on Vercel)

The live site is a static bundle served directly on Vercel — no per-request compute and no AWS in the request path. Snapshot the catalog and publish:

python scripts/build_static.py                 # snapshot DB + UI into dist/
( cd dist && npx vercel deploy --prod --yes )  # publish to conferenceagent.vercel.app

Search, sort, CSV export, and per-row .ics download all run client-side over the JSON snapshot. Cloudflare Pages (npx wrangler pages deploy dist) is an equivalent static host. See DEPLOY.md.

Legacy AWS path (no longer the deploy target). The web table can also run as a FastAPI app on AWS Lambda + RDS PostgreSQL via the SAM template in infra/ (pip install -e ".[web,deploy]"; cd infra && sam build && sam deploy --guided). It is kept as a dynamic-backend alternative — see DEPLOY.md.

To test locally:

python -m conference_agent.cli serve

then follow URL

License

MIT