Paper Pilot
Paper Pilot is a local MCP research server that lets an AI agent search literature, download and read full papers, inspect figures/tables, and produce evidence-backed research outputs.
Search six academic databases (Semantic Scholar, OpenAlex, Europe PMC, arXiv, Crossref, DOAJ) with year and open-access filters.
Find similar papers from a seed title/DOI and generate interactive citation/relatedness graph HTML for a topic.
Download/inspect open-access PDFs, extract full text into searchable chunks, and retrieve exact page text or rendered page images for figures and tables.
Use
deep_read_topicfor end-to-end search, download, extraction, evidence chunks, local PDF paths, optional page rendering, and optional PDF embedding.Run
research_topicend-to-end and optionally sync results to Zotero; list Zotero collections.Access optional LibGen and Sci-Hub tools in the schema for supplemental search/download (Sci-Hub requires
SCIHUB_ENABLED=true).Check current configuration and enabled integrations with
healthcheck.
Searches academic papers from arXiv as part of multi-source scholarly search, enabling AI agents to find and access research papers from this preprint repository.
Resolves Digital Object Identifiers (DOIs) for academic papers, enabling AI agents to locate and access specific research papers through DOI-based lookups and Sci-Hub integration.
Searches academic papers from Semantic Scholar as part of multi-source scholarly search, providing access to scholarly literature with citation data and metadata.
Syncs research papers, metadata, and reports to Zotero libraries, enabling AI agents to automatically organize and manage academic references in both local and web-based Zotero instances.
Paper Pilot
Let Codex and Claude find, download, and read full research papers.
Search six academic databases, read methods and results page by page, inspect figures, and get back to the decision you are trying to make. Paper Pilot is a local MCP server; your existing AI agent does the reasoning.
Install · Real example · Reading guide · Report a problem
Quick start
Install uv and Git if needed. Choose your client:
Codex
codex mcp add paper_pilot -- uvx --from git+https://github.com/aytzey/paper-pilot paper-pilotClaude Code
claude mcp add --scope user paper-pilot -- uvx --from git+https://github.com/aytzey/paper-pilot paper-pilotRestart your client, then ask:
Use Paper Pilot to find “Attention Is All You Need”. Read the full paper, inspect the architecture figure, and explain the method and its limitations with PDF page citations.
No account, API key, email setup, or Zotero installation is required by Paper Pilot. If you search a lot, add the two free keys in Search limits. Your AI client's own access and usage costs still apply. The first launch installs Python dependencies; academic APIs can rate-limit requests.
Claude Desktop or Cursor: use the copyable JSON configuration.
Try the tools without an AI client:
uvx --from git+https://github.com/aytzey/paper-pilot paper-pilot demo "retrieval augmented generation"The CLI downloads accessible papers, saves a reading pack, and opens a citation graph. It does not run an LLM or write a finished research conclusion. Results go into ~/.paper-pilot/ by default. These commands install from GitHub; a PyPI release is not required.
Related MCP server: Research MCP
See what reaches your agent
In a recorded stdio MCP check, Attention Is All You Need was delivered as 15 pages of text across two responses. All 39,498 extracted characters arrived in order. The original 2,215,244-byte PDF also arrived unchanged through both supported PDF transfer paths.
What was checked | Recorded result |
Whole-document text | Every batch joined to exactly match extraction from all PDF pages |
Original PDF | Returned bytes matched the downloaded file |
A multidisciplinary example | Two soil-pH papers: 37 pages, 96,163 characters, six responses |
A paper whose publisher blocks scripts | Read in full from the PubMed Central open-data copy |
A paywalled paper with an author manuscript in PubMed Central | Read in full; the response says which copy it is |
Response size | Default search, download and reading responses measured 4,300 to 7,400 tokens |
The last row matters more than it looks. Codex truncates tool output at roughly 10,000 tokens and Claude Code at 25,000. A response above that is cut by the client, and the agent is left with an abstract and the first pages without knowing it. Every default response is sized to arrive whole.
Coding agents can open files themselves, so why does the server send the text? Because it was tested. Given only file paths, Claude Code read the extracted text file and was a little faster; Codex ignored the file and asked the server for the text anyway, one call later than it needed to. So the server sends the text, lists every remaining batch up front so the rest can be fetched in parallel, and also returns text_path and pdf_path for agents that prefer their own tools. The walkthrough has the runs.
Read the walkthrough and reproduction steps or inspect the verification data. The transport check verifies delivery, not model comprehension. Separately, one Claude Code session and one Codex CLI session were each asked four questions about a 26-page paper: three answered only on pages 22 and 26, one only inside a figure. Both read to the end, rendered the figure page, and answered all four correctly.
Research a practical decision
Start with something you need to decide:
I want to build an inexpensive circuit to measure soil pH. Use Paper Pilot to investigate soil science, electrode materials, and electronics, starting with recent papers and following foundational references. Read the decisive papers in full. Compare cost, calibration, drift, and measurement conditions. Recommend an approach, cite the evidence, and propose the first experiment to validate it.
The server supplies research instructions alongside its tools: split the question across disciplines, continue the relevant searches, inspect the full papers, seek contrary evidence, and disclose gaps before recommending an approach. The full eight-step workflow is also an MCP prompt, research_decision (in Claude Code: /mcp__paper-pilot__research_decision).
Other useful starting points:
Your decision | Ask your agent |
Choose a RAG design | “Compare recent RAG approaches for my document set. Read the evaluation sections and explain which results transfer to my constraints.” |
Understand a disputed result | “Find papers supporting and challenging this claim. Compare their methods and conditions with page citations.” |
Inspect a paper you already have | “Read this PDF completely. Check the main figures and list the assumptions behind its conclusion.” |
Search covers Semantic Scholar, OpenAlex, arXiv, Crossref, Europe PMC, and DOAJ. To get the full text, download_paper tries arXiv, the PubMed Central open-data store, every PDF link those indexes know, Unpaywall locations, and finally PubMed Central looked up by DOI. Zotero sync is optional.
Not every paper has an open copy. In a check of 60 search results across five topics, 40 could be read in full and 20 could not: paywalled articles and dead links. For those, the tool returns the reason, and the agent is told to report the paper as unread instead of leaning on its abstract.
Search limits
Two of the six sources ration requests without a key. OpenAlex gives each network about 100 keyless searches a day, and Semantic Scholar's keyless pool is usually busy. When either runs out, search continues on the other sources and the response says which one was skipped and why. Both keys are free:
OPENALEX_API_KEY=... # https://openalex.org/settings/api-key
SEMANTIC_SCHOLAR_API_KEY=... # https://www.semanticscholar.org/product/apiSet them as environment variables of the MCP server (client setup). Downloading and reading papers does not need them.
How full-paper reading works
flowchart LR
Q[Your question] --> S[Search across disciplines]
S --> P[Download selected PDFs]
P --> R[Read text and inspect figures]
R --> A[Agent compares the evidence]Download:
download_papertakes a DOI, an arXiv ID, a PubMed Central ID, a PDF link, or just the exact title of a reference the agent wants to follow. It returns the first batch of text and the list of every batch still to read.Text:
read_pdf_textreturns consecutive pages, including pages longer than a single response. A batch holds up to 20,000 characters, so a typical paper takes two to four calls, and the remaining ones can be sent together.Figures and tables:
render_pdf_pagesreturns page images for visual inspection.Original file:
read_pdf_documentprovides the PDF when your agent needs it and the client supports that delivery path.Evidence trail: retain source, DOI, PDF page, access location, and extraction warnings.
The agent must follow every continuation to complete a text read. Scanned PDFs without a text layer are reported as unreadable; OCR is outside the project scope. Search results reflect returned provider pages, so exhaustive literature coverage is not guaranteed.
See the tool-level reading guide for exact calls, access rules, and limits.
MCP client setup
For Claude Desktop or Cursor, add this server to your MCP configuration:
{
"mcpServers": {
"paper-pilot": {
"command": "uvx",
"args": ["--from", "git+https://github.com/aytzey/paper-pilot", "paper-pilot"]
}
}
}Ready-to-copy files: Claude Desktop · Cursor · Codex.
On Windows, if the client cannot find uvx, use the executable's full path from where.exe uvx. Keep the arguments unchanged. Client setup covers configuration locations, local checkouts, updates, PDF capabilities, and optional Zotero setup.
Where Paper Pilot fits
Use it when you want your existing agent to research a question across disciplines and inspect the actual papers. If your main task is different, these projects are also worth a look:
Project | Main focus |
arXiv literature workflows, original LaTeX section reads, BibTeX, and topic watches | |
Working with an existing Zotero library through an AI assistant | |
A RAG system for answering questions from scientific documents with citations |
These are different starting points, not a quality ranking. Paper Pilot does not include its own LLM or require a vector database.
Contribute a real research case
Try a paper or decision you know well. If the agent misses a source, cannot retrieve a PDF, or loses part of a page, open an issue with the DOI, client, and failed step. A small reproducible case is particularly useful.
For code contributions, see CONTRIBUTING.md. The most useful work is reliable source access, reading correctness, and verified client instructions.
Architecture · Agent instructions · Codex guide · Claude guide · Optional integrations
MIT licensed. If this is useful for your next research task, star Paper Pilot to keep it handy. Sharing a case that worked helps other people decide whether to try it.
Available Tools
16 toolsdeep_read_topicA
Search, download, extract full text, and return evidence chunks plus local PDF paths for direct inspection.
By default the result is just [result_dict]: it carries every downloaded PDF's local path in
pdf_paths (and in deep_reads[].pdf_path / downloads[].pdf_path), so a client can open the
files when needed without any base64 in the payload. Opt in to inline content when your client
benefits: render_top_pages=True appends the top paper's relevant pages as images (vision models);
attach_top_pdf=True embeds its PDF as an application/pdf resource (Claude API style). include_scihub=True
adds a Sci-Hub fallback; write_graph=True also renders a citation graph (path in graph_path).
Inputs: topic (required), research_question, limit_per_source, related_limit, +21 more.
download_scihub_paperA
Download a paper PDF via Sci-Hub using its DOI. Returns local path and text preview. Requires SCIHUB_ENABLED=true.
Inputs: doi (required), topic_hint.
extract_local_pdf_textC
Extract full text from a local PDF, save a text sidecar, and return top matching chunks.
Inputs: pdf_path (required), title_hint, research_question, chunk_size_chars, +2 more.
find_similar_papersC
Find similar papers starting from a seed paper title or DOI.
Inputs: seed_title (required), seed_doi, limit, open_access_only.
get_pdf_page_textA
Return the exact extracted text of specific PDF pages (1-based) as plain JSON.
For fine-grained lookups over the wire (a single reference entry, a table, a footnote) without base64 and without filesystem/shell access. Use deep_read_topic first to get the pdf_path and the page mapping (deep_reads[*].chunk_manifest_path), then fetch the exact pages you need here.
Inputs: pdf_path (required), page_numbers (required).
graph_topicB
Search a topic and render an interactive citation/relatedness graph as a self-contained HTML file.
Nodes are papers (size scales with citation count, color by year); edges connect papers that share keywords or a venue, plus similarity links radiating from the top result. Returns the local HTML path. Open it in a browser to explore or screenshot the landscape.
Inputs: topic (required), limit_per_source, related_limit, from_year, +2 more.
healthcheckA
Return current configuration summary and enabled integrations.
Inputs: no named parameters.
inspect_libgen_itemC
Resolve LibGen mirror links, download a PDF when possible, and return a preview.
Inputs: mirror_1 (required), title (required), author, year, +3 more.
inspect_open_access_pdfC
Download an open-access PDF and return a local preview.
Inputs: pdf_url (required), filename_hint.
list_zotero_collectionsC
List Zotero collections visible to the configured web or local Zotero integration.
Inputs: query.
read_pdf_documentA
Return a downloaded PDF's local path so a client can open the file when it needs it.
By default nothing is base64-inlined: the result is [metadata, resource_link] where metadata carries pdf_path / doc_id / page_count / size_bytes. The PDF stays on disk; clients that read local files (e.g. Claude Code) or fetch the paperpilot://pdf/{doc_id} resource use it without payload bloat. Set embed_base64=True to inline the PDF as an application/pdf resource for clients that read PDFs that way (Claude API integrations), bounded by max_mb / max_pages.
Inputs: pdf_path (required), embed_base64, max_pages, max_mb.
render_pdf_pagesA
Render selected PDF pages to PNG and return them as images the model can see.
The result is [metadata_dict, image, image, ...]: the dict keeps pdf_path and
images (local PNG paths) for path-based clients, and the trailing image blocks
let a vision-capable model inspect figures, tables, and layout directly. Set
include_images=False to return only the paths.
Inputs: pdf_path (required), page_numbers (required), scale, include_images.
research_topicC
Run the end-to-end research workflow and optionally sync the result into Zotero. Set include_scihub=True to use Sci-Hub as a fallback for papers without open-access PDFs. Set write_graph=True to also render an interactive citation graph HTML (path returned as graph_path).
Inputs: topic (required), limit_per_source, related_limit, download_top_n, +14 more.
search_libgenC
Search LibGen mirrors for supplemental research material.
Inputs: query (required), search_type, limit, allowed_extensions.
search_literatureC
Search Semantic Scholar, OpenAlex, Europe PMC, arXiv, Crossref, and DOAJ for a topic.
Inputs: topic (required), limit_per_source, from_year, to_year, +1 more.
search_scihubB
Search Sci-Hub for papers by DOI, title, or keyword. Use search_type='doi', 'title', or 'keyword'. Requires SCIHUB_ENABLED=true.
Inputs: query (required), search_type, limit, check_availability.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 16 tools
Most tools have distinct purposes, but there is some overlap between deep_read_topic and research_topic, and between various search/download tools. Descriptions help differentiate them.
All tool names follow a consistent verb_noun pattern in snake_case, except for 'healthcheck' which is a single word but still clear. Overall very consistent.
16 tools is well-scoped for a research assistant server covering search, download, extraction, visualization, and Zotero integration. No extraneous tools.
Covers major workflows (search, download, extraction, graphing, Zotero) but lacks tools for managing Zotero items beyond collections or batch operations. Minor gaps.
Maintenance
Related MCP Connectors
Research paper search with real citations and reference formatting for AI assistants
Search arXiv/Semantic Scholar/OpenAlex + medical evidence (PubMed/Europe PMC) + LaTeX/PDF tools.
Search 340M+ academic papers — citation graphs, semantic similarity, and AI literature reviews.
Paper search (free), literature review packs, hallucinated-citation checks, arXiv digests, citations
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables AI assistants to search across multiple academic databases (PubMed, arXiv, bioRxiv, medRxiv, Semantic Scholar) through a unified interface. Supports advanced filtering, metadata retrieval, PDF downloads, and comprehensive research workflows with citation analysis.5-
- AlicenseNot gradedqualityDmaintenanceEnables LLMs to search, analyze, and summarize academic research papers in real-time from arXiv, Semantic Scholar, and PubMed. Provides automatic deduplication, citation analysis, and BibTeX generation across multiple research databases.22 npmMIT
- AlicenseNot gradedqualityDmaintenanceTurn any AI agent into an academic researcher that can search, read, cite, and write full literature reviews autonomously.14MIT
- AlicenseBqualityBmaintenanceAI-operable research workspace integrating Zotero, Obsidian, and NotebookLM. Search papers (arXiv/Semantic Scholar/PubMed/CrossRef), ingest into Zotero, sync per-paper notes to Obsidian, verify NotebookLM briefs. All three external tools optional.76117 PyPI70MIT