One endpoint gives any agent — Claude, Cursor, ChatGPT, or your own — structured access to every paper, overview, claim ledger, reproduction record, and bounty on AutoXiv, and lets it ask a paper — or the whole corpus — questions, register arXiv papers, and run the Alpha Machine pipeline. Read-only calls work with no signup; a free key unlocks writes, questions, and pipelines.
POST https://autoxiv.com/mcpGET https://autoxiv.com/mcp returns live server info, tiers, and limits — point an agent at it and it can self-onboard.Once connected, hand your agent one of these — each exercises a different slice of the server.
The corpus values in the first two tabs were fetched live when this page rendered; the pipeline tab shows the async call shapes. Then fire a call yourself, no key needed.
tools/call → list_recent { "limit": 10 }| Tier | Auth | Limits | Can do |
|---|---|---|---|
| Trial | none (per-IP) | small daily cap | all read tools — zero signup |
| Free key | Authorization: Bearer ak_live_… | 20 req/min · 1,000/day | reads + all write scopes below |
| Scope | Gates |
|---|---|
| write:library | save_paper, unsave_paper |
| write:review | request_code_review |
| write:chat | chat_with_cluster, ask_paper, ask_corpus |
| write:alpha | start_alpha_run |
| write:reproduce | start_reproduction |
| write:submit | register_arxiv_paper |
Create and revoke keys at /account/api-keys — the same page shows a usage dashboard: per-key daily call counts and the latest calls with tool, status, and latency. New keys carry every scope; a key created before a scope shipped (write:alpha, write:reproduce, write:submit) needs to be re-created to use the tools that scope gates.
start_alpha_run returns a run_id in under a second while the pipeline (typically 60–300s) runs server-side; poll get_alpha_run every ~20s until status is complete or failed. The run also appears in your /alpha-machine history with downloadable .py / .ipynb / .md deliverables. request_code_review → get_code_review follows the same pattern, and so does register_arxiv_paper → get_overview: the paper is published on return, its AI overview lands a little later. Alpha Machine output is research tooling, not investment advice.v0.9 — ask_corpus: one question across the whole corpus, answered only from the papers retrieved for it (semantic search plus a keyword pass, up to 10 papers’ overview fields and reproduction evidence), every claim cited by paper id, with the papers consulted, their evidence labels, and the ids cited. Same logic as the new /ask page; 20/day per key under write:chat; a question nothing matches costs nothing.v0.8 — Alpha Machine executes its own backtest scaffold: a Vercel Sandbox (python3.13, pandas + numpy, outbound network denied where the sandbox supports it — otherwise the package-registry allowlist, recorded on the run) runs the generated file against its synthetic data, with one automated repair when it crashes. get_alpha_run gains execution (ran / repaired / failed / skipped + reason, duration, parsed metrics) and the memo states whether the scaffold ran. Synthetic data: a run proves the file executes, not that the strategy works.v0.7 — the paper page’s two newest features, agent-callable: ask_paper (one grounded question per call, answered only from AutoXiv’s record of the paper, with page citations; 40/day per key under write:chat) and register_arxiv_paper (link any arXiv paper by id or URL with an optional repo, overview generated in the background, idempotent on re-registration; 10/day per key under the new write:submit scope). Both share their server logic with the web routes.v0.6 — verification release: typed claim ledger (claim.v1) with a blind second extraction pass and a database-enforced confirmed-only gate; get_paper_claims; server-captured run manifests; structured result channel with executor/judge separation; multi-seed replicates with Wilson / BCa intervals for stochastic claims.v0.5 — integrity release: demo evidence purged and filtered everywhere; match scores + claimed metrics are now server-computed (never model-self-reported); verdict taxonomy separates untestable/over-budget from failed; repo output fenced as untrusted in the reproduce agent; bounty settlement frozen pending adjudication; claim eligibility hardened (successful verdict, no self-claims, evidence must postdate the pot); score rubrics and match methodology published.v0.4 — reproduction over MCP (start_reproduction / get_reproduction_run, write:reproduce scope): the full replicate-a-paper loop is now agent-callable. Worked example sessions + live in-browser demo on this page.v0.3 — list_open_calls (OpenReview venue discovery) + the /conferences board, per-client install tabs with one-click Cursor/VS Code setup, example prompts, TradingView market view in Alpha Machine results, evaluation scorecards on paper pages.v0.2 — Alpha Machine over MCP (start_alpha_run / get_alpha_run, new write:alpha scope), list_bounties, usage dashboard on the keys page.v0.1 — read tools, library + review + cluster-chat writes, trial tier, per-key rate limits.