alphaxiv-py

alphaxiv-py is a Python SDK and CLI for alphaXiv.

This v1 release is intentionally public-first. It supports:

• API-key auth via ALPHAXIV_API_KEY or a locally saved alphaXiv API key
• optional browser-backed web login for assistant commands when API-key chat writes are restricted
• authenticated assistant chat for homepage and paper-scoped sessions
• homepage search suggestions for papers, topics, and organizations
• homepage feed/card retrieval with sort and filter support
• bare and versioned arXiv ID resolution
• paper metadata
• public paper comments
• paper full text
• AI overview/blog payloads
• overview status and translation availability
• similar-paper recommendations
• mentions/resources aggregation
• podcast transcript retrieval
• PDF URL lookup and download
• authenticated folder listing
• authenticated assistant URL metadata lookup
• authenticated paper votes plus comment create, reply, delete, and upvote mutations
• CLI context around a current paper and assistant session

Authenticated features use direct HTTP bearer auth against api.alphaxiv.org. The recommended setup is an alphaXiv API key exposed through ALPHAXIV_API_KEY or saved locally with alphaxiv auth set-api-key. If your account's API key can read assistant metadata but cannot start or reply to chats, run alphaxiv auth login-web; assistant commands prefer the saved web login when it is available. The client reads the current preferred model live from GET /users/v3 and writes changes through PATCH /users/v3/preferences; it does not claim to know the full current model catalog.

Installation

uv add alphaxiv-py

or

pip install alphaxiv-py

For local development:

uv pip install -e .

Optional browser support for alphaxiv auth login-web:

uv sync --extra browser
uv run playwright install chromium

Or from PyPI:

pip install "alphaxiv-py[browser]"
playwright install chromium

Recommended auth setup:

export ALPHAXIV_API_KEY="axv1_..."
alphaxiv assistant list

If you prefer to save the key locally:

alphaxiv auth set-api-key --api-key "$ALPHAXIV_API_KEY"

If assistant chat writes are restricted for your API key, save the alphaXiv web session too:

alphaxiv auth login-web

Use auth login-web as one-time setup for a persistent browser profile, not before every assistant command. After it succeeds, just use alphaxiv assistant ... normally and the CLI will try to refresh the saved web token automatically from ALPHAXIV_HOME/browser-profile.

If someone keeps rerunning auth login-web, the usual problem is operational rather than token-lifetime alone:

• ALPHAXIV_HOME changes between runs
• the environment is ephemeral and does not keep ~/.alphaxiv
• browser-profile is being deleted or recreated
• the underlying alphaXiv web session itself was signed out

Tell users to keep ALPHAXIV_HOME stable, keep browser-profile, and rerun auth login-web only when the real web session is gone or they moved to a new machine or container.

Users create API keys in the alphaXiv web app. This package does not create or manage API keys remotely.

CLI Quick Start

The CLI is resource-first. The top-level surface is grouped into auth, context, search, feed, paper, assistant, folders, guide, skill, and agent.

export ALPHAXIV_API_KEY="axv1_..."
alphaxiv auth set-api-key --api-key "$ALPHAXIV_API_KEY"
alphaxiv auth status
alphaxiv auth login-web
alphaxiv guide research
alphaxiv context use paper 2603.04379
alphaxiv context show
alphaxiv context show --json
alphaxiv search all "graph neural networks for molecules"
alphaxiv search papers "graph neural networks for molecules"
alphaxiv search papers "graph neural networks for molecules" --json
alphaxiv search organizations "graph neural networks for molecules"
alphaxiv search topics "graph neural networks for molecules"
alphaxiv assistant list
alphaxiv assistant list --json
alphaxiv assistant model
alphaxiv assistant url-metadata https://github.com/PKU-YuanGroup/Helios
alphaxiv assistant set-model "Claude 4.6 Sonnet"
alphaxiv assistant start "Find papers on agent frameworks"
alphaxiv assistant start --model "GPT 5.4" "Find papers on agent frameworks"
alphaxiv assistant reply "Focus on the most cited ones"
alphaxiv assistant history
alphaxiv feed filters
alphaxiv feed filters search "agentic"
alphaxiv feed filters search "agentic" --json
alphaxiv feed list --sort likes --limit 5
alphaxiv feed list --organization MIT --source twitter --limit 5
alphaxiv feed list --topic agentic-frameworks --organization Meta --limit 5
alphaxiv paper show
alphaxiv paper abstract
alphaxiv paper summary
alphaxiv paper summary --json
alphaxiv paper comments list
alphaxiv paper comments add --body "Helpful note"
alphaxiv paper comments reply <comment-id> --body "Thanks"
alphaxiv paper comments upvote <comment-id> --yes
alphaxiv paper comments delete <comment-id> --yes
alphaxiv paper similar --limit 5
alphaxiv paper text --page 1
alphaxiv paper vote --yes
alphaxiv paper view --yes
alphaxiv paper overview
alphaxiv paper overview --machine
alphaxiv paper overview-status
alphaxiv paper resources
alphaxiv paper resources --bibtex
alphaxiv paper resources --transcript
alphaxiv paper pdf url
alphaxiv paper pdf download ./helios.pdf
alphaxiv folders list --papers
alphaxiv skill install
alphaxiv skill status --scope all --json
alphaxiv agent show codex
alphaxiv context clear
alphaxiv auth clear
alphaxiv auth clear-web

--json is the stable machine-readable output mode for automation and agents. --raw stays available on selected commands when you want the backend-shaped payload for debugging. Assistant commands prefer the saved web login when it is available, and otherwise fall back to API-key auth.

Long-running assistant chats can become slower over time because assistant reply keeps extending the same remote session. If a chat becomes sluggish and you do not need the whole thread active, start a fresh chat with alphaxiv assistant start "..." and carry forward only the minimal recap or paper ids you still need. Use alphaxiv assistant history or alphaxiv assistant list to keep the older session available for reference. This is the practical mitigation when the slowdown is coming from accumulated chat state.

Python Quick Start

import asyncio
import os

from alphaxiv import AlphaXivClient


async def main() -> None:
    async with AlphaXivClient(api_key=os.environ["ALPHAXIV_API_KEY"]) as client:
        homepage = await client.search.homepage("graph neural networks for molecules")
        print(homepage.papers[0].paper_id, homepage.papers[0].title)

        papers = await client.search.papers("graph neural networks for molecules")
        organizations = await client.search.organizations("graph neural networks for molecules")
        topics = await client.search.closest_topics("graph neural networks for molecules")
        print(papers[0].paper_id, organizations[0].name, topics[0])

        filter_search = await client.explore.search_filters("agentic")
        print(filter_search.topics[:2], [item.name for item in filter_search.organizations[:2]])

        cards = await client.explore.feed(
            sort="most-stars",
            source="GitHub",
            topics=("agentic-frameworks",),
            limit=3,
        )
        print(cards[0].paper_id, cards[0].upvotes)

        paper = await client.papers.get("2603.04379")
        print(paper.version.title)

        comments = await client.papers.comments("2603.04379")
        print(comments[0].body if comments else "no comments")

        full_text = await client.papers.full_text("2603.04379")
        print(full_text.pages[0].text[:200])

        overview = await client.papers.overview("2603.04379")
        print(overview.summary.summary if overview.summary else overview.title)

        status = await client.papers.overview_status("2603.04379")
        print(status.state, sorted(status.translations))

        transcript = await client.papers.transcript("2603.04379")
        print(transcript.lines[0].speaker, transcript.lines[0].line[:80])

        similar = await client.papers.similar("2603.04379", limit=3)
        print([card.paper_id for card in similar])

        sessions = await client.assistant.list()
        print(sessions[0].id, sessions[0].title)

        metadata = await client.assistant.url_metadata("https://github.com/PKU-YuanGroup/Helios")
        print(metadata.title)

        folders = await client.folders.list()
        print(folders[0].name if folders else "no folders")

        run = await client.assistant.ask("Find papers on agent frameworks")
        print(run.session_id, run.output_text[:120])


asyncio.run(main())

AlphaXivClient.from_saved_api_key() loads ALPHAXIV_API_KEY first and then falls back to ~/.alphaxiv/api-key.json. AlphaXivClient.from_saved_browser_auth() reuses the auth captured by alphaxiv auth login-web. AlphaXivClient.from_saved_auth(prefer_browser=True) prefers the saved web login and falls back to the saved API key.

Agent Integrations

This repo ships agent-facing guidance in two forms:

• built-in workflow guides through alphaxiv guide ...
• packaged integrations installable through alphaxiv skill ...

The repo-root Codex skill is also published for discovery:

• Skill entry point
• Skill workflows
• Skill command map

Install the packaged integrations with the CLI:

alphaxiv skill install
alphaxiv skill status --scope all
alphaxiv skill show --target source
alphaxiv agent show codex

Supported targets in this release:

• Codex
• Claude Code
• OpenCode

skill install defaults to user scope and installs all supported targets. Use --scope project to install into the current project instead.

LLM Agent Tips

If an agent is driving the CLI, the intended flow is:

1. use search for keyword lookup
2. use feed for recent or ranked discovery
3. use paper to inspect one paper
4. use assistant after retrieval for synthesis or follow-up reasoning
5. use context when several commands target the same paper or chat

Output modes:

• use --json for stable machine-readable command output
• use --raw only when you need the backend-shaped payload for debugging

High-confusion distinctions:

• paper abstract = original abstract
• paper summary = short AI digest
• paper overview = long AI write-up
• paper text = readable text extracted from the PDF
• paper pdf download = actual PDF file

Docs

• CLI reference
• Python API
• Development
• Releasing
• API inventory