MCP service · curated database · soon-open-source extraction toolchain. Built for serious CV/ML researchers. Plug into Claude Code; survey a field in 5 minutes, drill into any paragraph in any paper, compare methods and experimental data across 100+ papers in one call.
Venues covered: CVPR · ICCV · ECCV · ICLR · ICML · NeurIPS · AAAI · ACL · EMNLP · NAACL · IJCAI · WACV · BMVC · 3DV · SIGGRAPH · TPAMI · IJCV · Refreshed weekly.
Designed for the way researchers actually work in 2026: an LLM agent at your fingertips, doing the reading. We give that agent the knowledge base — pre-distilled, structured, cite-locked.
One MCP connection grants your agent search, quote, compare_methods, find_baselines, survey, trends, narrative_threads, get_figure, bibtex and 9 more — designed per Anthropic's tool-design guidelines with defer_loading.
59K papers from CV/ML/NLP top venues, distilled with GPT-5.5 and Opus-4.7 against open schema v1.3 — 7-class contribution taxonomy + 6-field narrative arc + dataset/metric grids + figure metadata. Hybrid retrieval: BGE-1024 dense + BM25 + RRF + Qwen3 cross-encoder reranking.
Schema spec, distillation prompts (Opus / GPT-vision), Marker/PyMuPDF figure extraction, LanceDB + SQLite FTS5 index builder, citation graph resolver, FastMCP server — all releasing Apache 2.0. Cards as CC-BY-SA dataset (Hugging Face). Ingest your lab's private papers, expose to your team.
acceptpaper isn't a scrape. It's a four-stage pipeline — and every stage is regression-tested against a held-out evaluation harness, so the numbers above stay honest.
64,292 top-venue PDFs crawled (CVPR · ICCV · ECCV · ICLR · ICML · NeurIPS · ACL · EMNLP …) → 59,631 in the live index. 99.6% of the corpus is 2024–2026 — the frontier, not a snapshot a model memorized two years ago.
GPT-5.5 vision reads each PDF into 30+ structured fields — 7-class contribution taxonomy, 6-field narrative arc, eval grids with numeric values + SOTA flags, baselines, key modules, figure metadata. QA'd corpus-wide: 98–100% field completeness, ~84% claim-faithful, hallucinations in the single digits per 100 cards.
1.15M paragraphs (0.03% garbled), 16,864 hero figures, a 997K-edge citation graph, 106K authors, 940K facets — in LanceDB + SQLite FTS5. Independently verified read-only: 325 live tool calls, 0 errors; SQLite ↔ FTS5 ↔ LanceDB counts aligned; every vector 1024-dim, norm≈1, no NaN.
BGE-1024 dense + BM25 + RRF fusion + Qwen3-0.6B cross-encoder rerank — ~0.5s p95, cite-locked to the source paragraph. Relevance@3 = 100% on a blind audit — it matches intent, not just keywords.
crawl → distill → index → retrieve · every stage measured, not assumed
Optimized for agent consumption per Anthropic best practices: every tool has explicit "when to use" + "when NOT to use" clauses, deferred tools auto-discovered via discover_tools.
Connect from Claude Code (or any MCP-compatible agent). Get a free key from your dashboard, then replace YOUR_KEY below.
Other MCP clients (Codex, custom LangChain agent): plain HTTP + JSON-RPC over streamable-http transport. Auth via X-API-Key header or Authorization: Bearer.
# in your terminal claude mcp add \ --transport http \ -s user \ litscan-rag \ https://mcp.acceptpaper.com/mcp \ -H "X-API-Key: YOUR_KEY"
# drops both skills into ~/.claude/skills/ curl -fsSL https://acceptpaper.com/skills/install.sh | sh
# in a Claude Code session > Survey sparse-view 3D human reconstruction > papers from 2024–2026. Identify dominant > method families, anchor papers, key > benchmarks, and emerging directions. # Claude internally: # 1. calls survey(topic, k=50) → field map # 2. trends(topic) → temporal signal # 3. quote(...) on contested claims # 4. find_baselines(top-3) for comparison # → returns synthesized answer with cites
Existing tools were designed for human eyes on a webpage. acceptpaper was designed for an LLM agent acting on your behalf.
| Capability | acceptpaper | PaperQA2 | Elicit | OpenScholar |
|---|---|---|---|---|
| MCP-native (agent interface) | ✓ | — | — | — |
| Pre-distilled structured schema | v1.3 (7-class + 6-field) | flat text | semi-struct | flat |
| Paragraph-level retrieval | 1.15M chunks | ✓ | ✓ | ✓ |
| Figure retrieval (hero JPG) | 16K | v3 only | — | — |
| Citation graph (resolved) | 100K edges | tool | partial | via S2 |
| Venue-tier-aware ranking | ✓ | — | — | — |
| Cost per query (retrieval) | $0 | $0.05–2 | $0.01–.50 | ~$0.10 |
| Domain focus | CV/ML top-tier | biology-heavy | broad | broad |
| Open source license | Apache 2.0 (in prep) | ✓ Apache | closed | ✓ Apache |
| Self-hostable for lab corpus | soon | ✓ | — | partial |
We ran an agent on acceptpaper's MCP + the research-mentor skill against the same agent on memory alone, across 18 frontier CV/ML research tasks. We don't trust LLM-as-judge here — we measured κ=0.20, and judges even flag real 2025–2026 papers as fabricated. So we count something objective: verifiable citations.
Citations per task: native 0 → plain RAG 2–3 → acceptpaper (graph tools + agentic workflow) 6–13. Two ready-made skills drive it: research-mentor (critique your idea) and research-explorer (map a field's open problems).
Objective citation audit, not vibes — what makes research trustworthy is checkable sources. We broke our own ruler first: LLM-as-judge was too unreliable (κ=0.20) to grade this, so we count real citations instead.
A paper is more than text. Our distillation produces 30+ structured fields per paper — 7-class contribution taxonomy, 6-field narrative arc, eval datasets with numeric values + SOTA flags, baseline comparison strings, key modules, hero figure metadata. Agents read exactly the slice they need without re-parsing PDFs every query.
// excerpt of one L2 card { "source_id": "cvf:CVPR.2026:1024", "title": "DiHuR: Diffusion-Guided Generalizable Human Reconstruction", "contribution_type": { "primary": "combination", "secondary": ["empirical_study", "ablation_heavy"] }, "narrative": { "problem": "Reconstructing detailed 3D humans from ~3 sparse cameras…", "insight": "SMPL vertices map to consistent semantic regions…", "evidence": "CD 1.117 vs GP-NeRF 3.876 on THuman; gains hold on ZJU-MoCap" }, "eval_data": [ {"name":"THuman","metric":"Chamfer Distance","value":1.117,"is_sota":true}, // + 9 more entries ], "compares_with": ["NeuS", "SparseNeuS", "PIFuHD", "GP-NeRF", "SIFU"], "hero_figure": {"page":1, "caption":"Given 3 views with minimal overlap…"} }
Not prose — structured cards pulled live via deep(). Each carries exact SOTA numbers, the paper's own verbatim limitations, and its baseline graph. Three real examples across domains:
Segment Anything in Images and Videos
DAVIS-2017 90.7 J&F (SOTA); SA-V 77.9 vs Cutie 61.3; image seg at 130 FPS — 6× faster than SAM.
None stated — the paper lists no limitations, so the card leaves it empty. We record absence; we never invent one.
SAM · XMem++ · Cutie · DEVA · JointFormer · STCN
Open-world monocular depth foundation model
DA-2K 97.4% (SOTA) vs V1 88.5%; Ours-Small 60 ms / 25M params vs Marigold 5.2 s / 948M — >10× faster.
“The largest DINOv2-G teacher is resource-intensive, with 1.3B parameters, and is not suitable for many practical applications.”
MiDaS · Marigold · Geowizard · ZoeDepth · Metric3D · DPT
Real-time open-vocabulary object detection
LVIS zero-shot 35.4 AP @ 52 FPS (SOTA) — vs DetCLIP-T 34.4 AP @ 2.3 FPS (≈20× faster).
“fine-tuning the CLIP text encoder on Objects365 can degrade generalization because Objects365 contains only 365 categories and lacks abundant textual information.”
GLIP · Grounding DINO · DetCLIP · ViLD · RegionCLIP · Detic
Pulled from the live index via deep() — every number and quote is in the source paper, not generated.
The hosted service stays free for academic use. The toolchain ships open-source so any lab can self-host their own corpus.
Free API key for academic / research use. We ask only that you tell us briefly who you are and what you're working on — to prioritize features and venue coverage.
Free for academic use · instant self-serve key · 1000 queries/day