confos 0.1.0

Conference intelligence in your terminal — for humans and agents.

confos

Conference intelligence in your terminal — for humans and agents.

confos turns AI/ML conferences (NeurIPS, ICLR, ICML, COLM, …) into a local, queryable knowledge base. Ingest a venue once from OpenReview, then search papers, find the people working on a topic, see what's trending, visualize the landscape, and export agent-ready context — all locally, all scriptable, all with provenance back to the source.

It is built for terminals, shell scripts, CI, and coding agents (Claude Code, Codex, or any agent with shell access). Every command speaks stable JSON, and after a one-time ingest every query is local and offline — answers come back in milliseconds, and unlike asking an LLM, every number is real and traceable, not guessed.

confos ingest neurips-2025
confos papers search "long-running agents" --venue neurips-2025
confos authors find --topic "agent memory" --venue neurips-2025   # who works on this?
confos trends topic "evals" --venues neurips-2024,neurips-2025      # what's rising?
confos viz topics --venue neurips-2025                              # see the landscape
confos export context --topic "agent evals" --venue neurips-2025 --json

---

Why confos exists

Conference data is public but not usable. OpenReview and conference sites are browsing surfaces — fine for clicking through one paper at a time, useless for "show me everyone working on X, ranked," "how did this topic change year over year," or "give my agent everything it needs to plan a literature review." PaperCopilot has great dashboards but they live on the web, you can't script them, and an agent can't drive them.

confos owns the missing layer: a local, agent-native, composable index with real analysis (trends, people discovery, aggregates, graphs) and honest provenance. The differentiator is not "search papers" — it's that an agent can drive the whole surface to do real work, offline, repeatably.

What you can do

• Search papers by full text (title, abstract, keywords) with ranking and filters.
• Find people — rank the authors actually working on a topic, with their papers, affiliations, and a relevance explanation.
• Explore authors, organizations, and topics; follow related papers.
• Analyze — aggregate stats (topics, orgs, countries) with visible data-quality.
• Trends — how a topic moves across years / venues; compare two venues head-to-head.
• Visualize — terminal charts, plus exportable HTML/Mermaid topic & co-authorship graphs.
• Export — context packs (JSON/Markdown) purpose-built for agents; CSV/JSONL dumps.

Who it's for

• Researchers & enthusiasts preparing for a conference or surveying a field.
• Builders / founders tracking what labs and people are doing in an area.
• Agents — confos is a first-class tool for Claude Code / Codex: stable --json, clean stdout/stderr separation, a bundled skill, and provenance so the agent never has to hallucinate a statistic.

Install

Requires Python 3.12+ and uv. confos isn't on PyPI yet, so install it from source:

git clone https://github.com/RRaphaell/confos
cd confos
uv tool install .               # install the `confos` command globally
# or: uvx --from . confos ...   # run without installing
# or: uv sync && uv run confos  # to develop on it

First run

confos init                       # one-time: create the local store at ~/.confos
confos ingest neurips-2025        # pull the venue (network; ~4–5k papers, a few minutes)
confos papers search "agents" --venue neurips-2025
confos authors find --topic "agent memory" --venue neurips-2025

confos venues search "ICLR 2026" (or confos venues aliases) shows what you can ingest. After ingest, everything is local and offline. See docs/PRODUCT.md for the full tour and docs/ARCHITECTURE.md for how it works.

Documentation

Doc	What's in it
docs/PRODUCT.md	Goal, the wedge, who uses it, worked examples for every capability
docs/ARCHITECTURE.md	System design, ASCII diagrams, components, data model, stack
docs/CLI_CONTRACT.md	Full command tree, flags, output contract, exit codes, safety
docs/RANKING.md	How authors find ranks people + how --topic matching works
docs/SCHEMAS.md	Stable --json output shapes (the agent/script contract)
docs/BUILD_PLAN.md	How the project is built: phases, standards, research/notes discipline, testing, validation
docs/DECISIONS.md	Decisions + assumptions log (lightweight ADR)
docs/PROGRESS.md	Live build progress + session log
docs/REFERENCES.md	Lessons taken from ft, birdclaw, gogcli, create-cli
AGENTS.md	How an agent should drive confos
CONTRIBUTING.md	Setup, the gate, layering rules, conventions

Status

v0.1.0. The full v1 surface — ingest · search · people · orgs · stats · trends · viz · export · context packs · agent skill — is implemented, tested, and verified end-to-end against live OpenReview (scripts/live-test.sh). OpenReview is the only source in v1; more adapters, semantic search, an LLM ask, and an MCP server are designed-for-later (the seams exist, the code doesn't — see docs/PRODUCT.md §8). Built in public-grade phases; see docs/BUILD_PLAN.md and docs/PROGRESS.md.

Contributing

Contributions welcome — see CONTRIBUTING.md for setup, the gate, and the layering rules. The opt-in scripts/live-test.sh exercises the real API before a release.

License

MIT — see LICENSE.