Plenty of sites let a human search papers. AIConfPaper is built so your LLM agent can run the whole literature loop on its own: find accepted papers from top AI conferences by meaning, save them to a collection, pull each paper's LaTeX source and code, then develop new ideas or set up a quick experiment. One REST API, or a single pasted prompt. What your agent can do Search by meaning across ~190k accepted papers, then create a collection and add the best hits, saved to your account so you can open it in the UI later. Pull the LaTeX source of every paper in a collection, analyze the methods, and propose concrete ideas that improve on their shared limitations. Follow each paper's code link, clone the repositories, and set up a minimal experiment environment to try the methods. Give one of these to your agent (Claude Code, Cursor, and similar). It reads this page, works out the API calls, and runs the whole flow. Replace YOUR_KEY with your key from Account . Notes & limits for agents links.arxiv_src comes from /papers/ {id} (it is always null in search results). It can be null: retry once, but some papers have no source link at all even when an arXiv version exists elsewhere, so fall back to another paper rather than retrying forever. The LaTeX source is a gzip-compressed tar ( arXiv-<id>vN.tar.gz ); tar xzf extracts it. The main .tex may have a generic name (e.g. an unrenamed template). links.code is present for roughly a third to half of papers and is best-effort. Sanity-check the cloned repo's README/title against the paper before trusting it, since a link can occasionally point to a project page or a different project. To save a specific paper you already know by title: search the exact title ( /search?q=<the title>&limit=3 ), confirm the top result's title matches, then add its paper_id to a collection. For an arXiv paper outside the index, register it first via POST /papers/external . Limits: 120 requests per rolling minute per client. Every /api/v1 response carries X-RateLimit-Limit / X-RateLimit-Remaining ; a 429 carries Retry-After (seconds) — honor it instead of guessing a backoff. POST /papers/external additionally shares a global arXiv lookup queue: a new id can take a few seconds and may return 503 (retry shortly). Collection caps: 50 folders, 1000 papers per folder, 5000 papers total. Errors are RFC 9457 application/problem+json with a detail recovery hint. Prefer the batch endpoints: GET /papers?ids=… (up to 50 full records in one call) beats looping GET /papers/ {id} , and GET /collections/ {id} /export?format=bibtex returns a whole collection as one ready-to-use references.bib . Prefer tools over raw HTTP? Connect any MCP client (Claude Code, Claude Desktop, Cursor, and similar) to the Streamable HTTP endpoint at https://aiconfpaper.com/mcp . It exposes the read side as typed tools: search_papers , get_paper , batch_get_papers , find_similar_papers , list_facets . Collections (create, add, share, export) stay on the REST API below. Auth is the same key, sent as an Authorization header. Every request needs a free key from Account , sent as Authorization: Bearer acp_… . Errors are RFC 9457 problem+json with a recovery hint. Rank papers by semantic similarity to q . Filters: venue , year_min , year_max , sort . Full record: abstract, BibTeX, citation count, and links ( links.pdf , links.code , links.arxiv_src for the LaTeX source bundle). Also accepts arxiv:<id> ids. Batch detail: full records for up to 50 comma-separated ids in one call (unknown ids come back in missing ).

Nearest-neighbor papers in embedding space. Filters: venue , year_min , year_max . Valid venue / tier values and the year range. Create a collection. Body: {"name":"…"} . Returns its id . Add papers. Body: {"paper_ids":["…"]} — ids from /search , or arxiv:<id> ids from /papers/external . Idempotent; unknown ids are counted in skipped . Register an arXiv paper that is not in the index. Body: {"arxiv":"2107.02192"} (id or arxiv.org URL). Returns paper_id arxiv:<id> to use with add-papers. Remove papers (symmetric with add). Body: {"paper_ids":["…"]} . Idempotent; no more delete-and-rebuild to prune a collection. Rename and/or describe. Body: {"name":"…","description":"…"} (either field; empty description clears it). Toggle the public read-only link. Body: {"public":true} . Returns the /c/<slug> URL to hand to a collaborator (view / follow / copy, no edit). Toggle an EDITABLE invite link. Body: {"enabled":true} . Returns a /collab/<slug> URL; a logged-in user who opens it joins as an editor. The collection then appears in THEIR GET /collections ( owned: false ) and they can read / add / remove / export it with their own key. List who joined the collaboration invite (owner-only; includes emails). Revoke a collaborator's edit access (idempotent). One-call export of the whole collection: bibtex (a ready references.bib ), csv , or json . List your collections — your own and ones shared with you for co-editing ( owned: false ). List the papers in one of your collections (includes registered arXiv papers). Delete a collection you created (its papers go to trash). Lets an agent clean up after itself. Worked example: search to saved collection Machine-readable spec Point a tool at /openapi.json (OpenAPI 3.1) or /llms.txt (agent quickstart).