claude-paper-workflow

# Tool List > A Claude Desktop workflow for paper research, integrating arXiv MCP and PapersFlow MCP for paper search, PDF download, Chinese reading, citation verification, and literature review generation. A Claude Desktop workflow for Chinese scientific research reading, literature review, citation verification, and paper tracking. The core idea is to use two MCP tools separately: - `arXiv MCP Server`: responsible for arXiv paper search, PDF download, English paper to Chinese Markdown, and WeChat public account draft. - `PapersFlow MCP`: responsible for cross-database literature search, citation verification, related paper discovery, and citation network exploration. > Verification date: 2026-04-23. MCP client and third-party services update quickly. It is recommended to run the installation command again before official release to check. ## What is this / What is not This repository is: - A paper workflow description for `Claude Desktop`. - A minimal practice that integrates local `arXiv MCP` and remote `PapersFlow MCP`. - A Chinese explanation template suitable for direct posting on GitHub. This repository is not: - A complete `Claude Code CLI` tutorial. - A "large and complete" navigation page that stuffs all paper tools into it. - An endorsement of third-party MCP Server functionality. Formal research still requires verification from the original text. ## Applicable Scenarios - Track new papers in a certain direction over the past 1 to 6 months. - Quickly convert English arXiv PDF to Chinese reading notes. - Conduct systematic literature reviews, organize research contexts, representative papers, and blank points. - Verify whether references exist, reducing the risk of hallucinating citations. - Rewrite paper reading notes into public accounts, group meeting reports, or project research materials. ## Support Matrix | Combination | Recommended Method | Status | Description | | --- | --- | --- | --- | | Claude Desktop + `arXiv MCP Server` | Local MCP configuration file or local extension | Recommended | Suitable for local PDF download, conversion to Chinese Markdown, and WeChat draft generation | | Claude Desktop + `PapersFlow MCP` | Custom remote Connector UI | Recommended | Suitable for literature search, citation verification, citation network; not modifying local JSON | | Claude Code + `PapersFlow MCP` | `claude mcp add --transport http ...` | Optional | Command line usage, do not mix with Claude Desktop | ## Quick Start on Windows If you just want to get started with one version, follow this path: 1. Install Claude Desktop. 2. Install Node.js and confirm `node -v` and `npm -v` are available. 3. Create a new directory `C:\Users\YourUsername\Documents\ClaudePapers`. 4. Add `arxiv-mcp-server` to the local MCP configuration in Claude Desktop. 5. Add `https://doxa.papersflow.ai/mcp` to the remote Connector UI in Claude Desktop. 6. Restart Claude Desktop and open a new conversation to confirm that both tools can be called. ## Overall Process ```mermaid flowchart TD A["Research Question / Keywords / Time Range"] --> B["Claude Desktop"] B --> C["arXiv MCP Server"] B --> D["PapersFlow MCP"] C --> C1["arXiv Search"] C --> C2["Download PDF"] C --> C3["Chinese Markdown Interpretation"] C --> C4["WeChat Article Draft"] D --> D1["Cross-Database Retrieval"] D --> D2["Related Papers"] D --> D3["Citation Network"] D --> D4["Citation Verification"] C3 --> E["Reading Notes"] D1 --> F["Candidate Paper Pool"] D2 --> F D3 --> G["Research Context Diagram"] D4 --> H["Verifiable References"] E --> I["Review / Report / Public Account"] F --> I G --> I H --> I ``` ## Tool Division of Labor | Tool | Best for | Not recommended for | | --- | --- | --- | | `arXiv MCP Server` | Preprint search, PDF download, Chinese reading, WeChat public account draft | Not suitable for cross-database review and citation verification alone | | `PapersFlow MCP` | OpenAlex / Semantic Scholar / arXiv academic search, related papers, citation network, citation verification | Not responsible for downloading PDFs and converting them into complete Chinese notes | | Claude Desktop | Conversational scheduling, comprehensive judgment, report generation | Should not generate unverified references from memory | ## Preparation Before Installation 1. Install the latest version of Claude Desktop. 2. Prepare Node.js 18 or higher. On Windows, it is recommended to confirm `node -v` and `npm -v` are available. 3. Create a separate directory for paper output, such as `C:\Users\YourUsername\Documents\ClaudePapers`. 4. If using the Chinese interpretation function of `arXiv MCP Server`, you need to apply for a `SiliconFlow API Key`. 5. Do not submit real API Keys, downloaded PDFs, generated paper notes to public GitHub repositories. ## 1. Configure arXiv MCP Server `arXiv MCP Server` is a local stdio MCP Server, suitable for putting into Claude Desktop's local MCP configuration. ### Recommended Configuration: Claude Desktop Local MCP Claude Desktop supports installing local MCP through desktop extensions or using local configuration files. New users prioritize looking at Claude Desktop's `Settings > Extensions` and `Settings > Developer`; if you want to configure manually, the Windows path is usually: ```powershell $env:APPDATA\Claude\claude_desktop_config.json ``` Add the following content to `claude_desktop_config.json`. If the file already exists, only merge the new items in `mcpServers`, do not overwrite the original server. ```json { "mcpServers": { "arxiv-mcp-server": { "command": "cmd", "args": ["/c", "npx", "-y", "@langgpt/arxiv-mcp-server@latest"], "env": { "SILICONFLOW_API_KEY": "replace_with_your_siliconflow_key", "WORK_DIR": "C:\\Users\\YOUR_NAME\\Documents\\ClaudePapers" } } } } ``` macOS / Linux can usually be written as: ```json { "mcpServers": { "arxiv-mcp-server": { "command": "npx", "args": ["-y", "@langgpt/arxiv-mcp-server@latest"], "env": { "SILICONFLOW_API_KEY": "replace_with_your_siliconflow_key", "WORK_DIR": "/Users/YOUR_NAME/Documents/ClaudePapers" } } } } ``` After configuration, restart Claude Desktop and confirm that `arxiv-mcp-server` is connected in the chat box's `+` or `Connectors`. ### Available Tools | Tool Name | Purpose | | --- | --- | | `search_arxiv` | Search arXiv papers by keyword | | `download_arxiv_pdf` | Download PDF by arXiv ID or URL | | `parse_pdf_to_markdown` | Parse PDF into Chinese Markdown | | `convert_to_wechat_article` | Generate WeChat public account article draft | | `process_arxiv_paper` | One-stop process of downloading, parsing, and optionally generating WeChat draft | | `clear_workdir` | Clear work directory, use with caution | ## 2. Configure PapersFlow MCP `PapersFlow MCP` is a remote HTTP MCP Server, entry address: ```text https://doxa.papersflow.ai/mcp ``` ### Claude Desktop / claude.ai: Recommended Custom Connector For remote MCP, the recommended path for Claude Desktop and claude.ai is `Customize > Connectors > Add custom connector`, fill in the above MCP URL. Note that remote connectors are initiated by Anthropic cloud to connect MCP URL, not the same as local MCP in `claude_desktop_config.json`. If your Claude account or organization does not have access to custom connectors, you can use it in Claude Code, Codex, Cursor, Gemini CLI and other clients that support remote HTTP MCP. If Claude pops up a login or authorization page, complete it according to the process. Do not mix `ANTHROPIC_API_KEY`, `SiliconFlow API Key`, and any third-party tokens. ### Claude Code: Command Line Configuration The new Claude Code official documentation recommends using `http` transport for remote MCP: ```powershell claude mcp add --transport http papersflow https://doxa.papersflow.ai/mcp ``` If the PapersFlow old tutorial you refer to uses `--transport streamable-http`, and your Claude Code CLI still supports this writing method, you can also follow its tutorial. Both are connected to the same remote MCP URL. ### Available Public Tools | Tool Name | Purpose | | --- | --- | | `search` | General academic search | | `fetch` | Get paper metadata by DOI, title or URL | | `search_literature` | Literature search with filtering conditions | | `verify_citation` | Standardize and verify citations | | `find_related_papers` | Find related papers | | `get_citation_graph` | Build citation network of a paper | | `get_paper_neighbors` | Find citation network neighbors | | `expand_citation_graph` | Multi-hop expand citation network | ## 3. Recommended Paper Workflow ### 1. Narrow Down the Problem Don't just say "help me find diffusion model papers". Let Claude break down the topic into executable search protocols: ```text I want to study "efficient reasoning of diffusion models in video generation". Please help me break down this topic into: 1. Keywords and synonyms; 2. Suitable search formulas for arXiv / OpenAlex / Semantic Scholar; 3. Time range, default 2024-01 to present; 4. Inclusion and exclusion criteria; 5. Final output table fields. ``` ### 2. Build Candidate Paper Pool with PapersFlow ```text Please call PapersFlow MCP to conduct systematic literature search around "large language model reasoning ability". Requirements: 1. Prioritize papers after 2023; 2. Each result provides title, author, year, venue/arXiv, DOI/arXiv ID, citation count, and why it's relevant; 3. Classify by "reasoning dataset", "reasoning process supervision", "tool-enhanced reasoning", "search/verifier", and "evaluation benchmark"; 4. Return 30 candidate papers, do not fabricate missing fields. ``` ### 3. Deep Read Core Papers with arXiv MCP ```text Please call arXiv MCP to process this paper: [paste arXiv link or ID]. Output Chinese Markdown reading notes, structured as follows: 1. Problem solved by the paper; 2. Core method diagram; 3. Real innovation points compared to previous work; 4. Experimental setup and key metrics; 5. Failure cases and limitations; 6. 5 related papers I should continue reading. ``` ### 4. Use PapersFlow for Citation Network and Blank Points ```text Please use the 3 core papers as seeds to call PapersFlow MCP to expand one-hop and two-hop citation networks. Please output: 1. Key predecessor papers for each seed paper; 2. Basic works commonly cited by multiple papers; 3. Potential papers that have appeared in recent years but have low citations; 4. Research gaps, must correspond to specific evidence, do not talk vaguely. ``` ### 5. Verify Citations Before Writing Review ```text Below is the list of references I plan to put into the review. Please call PapersFlow's verify_citation to verify each citation: 1. Whether it exists; 2. Whether title, author, year, and venue match; 3. Whether DOI/arXiv ID can be parsed; 4. Mark as verified / ambiguous / not_found; 5. Provide revision suggestions for ambiguous and not_found. ``` ### 6. Output Final Review ```text Please write a Chinese literature review based on verified papers. Requirements: 1. Do not cite unverified papers; 2. Each conclusion is followed by corresponding papers; 3. Organized by timeline and method line; 4. List "existing consensus", "still controversial", and "new directions"; 5. Finally, provide BibTeX or DOI/arXiv ID list. ``` ## 4. Output Template It is recommended to let Claude output using this table, which is convenient for subsequent copying into Excel, Notion, or review drafts. | Field | Description | | --- | --- | | `paper_id` | DOI, arXiv ID, or PapersFlow return ID | | `title` | Paper title | | `year` | Year | | `venue` | Conference, journal, or arXiv | | `task` | What task is solved | | `method` | Method keywords | | `evidence` | Experimental or conclusion evidence | | `limitation` | Limitations | | `why_read` | Why it's worth reading | | `citation_status` | `verified` / `ambiguous` / `not_found` | ## 5. Verify Connection After configuration, do not immediately write a review. Perform minimal verification first. ### Verification 1: Is arXiv MCP Available? ```text Please call arXiv MCP to search for papers related to "test-time scaling reasoning" and return 3 results. ``` Expected result: Claude clearly displays `search_arxiv` or equivalent tool and returns arXiv results. ### Verification 2: Is PapersFlow Available? ```text Please call PapersFlow to search for "diffusion models video generation efficiency" and return only 5 papers. ``` Expected result: Claude clearly displays PapersFlow's search tool, not directly answering from memory. ### Verification 3: Is Citation Verification Available? ```text Please call PapersFlow verify_citation to verify "Attention Is All You Need, 2017, Vaswani et al.". ``` Expected result: Return standardized literature information and verification status. ## 6. Common Pits - Do not directly use `claude mcp add ...` as Claude Desktop's JSON configuration. It is mainly a command for Claude Code CLI. - Do not write an additional `papersflow` local JSON configuration for Claude Desktop. Remote Connector is not connected like this. - Windows Claude Desktop startup `npx` sometimes cannot find executable files. Prioritize using `command: "cmd"` and `args: ["/c", "npx", ...]`. - `WORK_DIR` must be written to a directory you have permission to access. The backslash in the path must be written as `\\` in JSON. - `clear_workdir` will clear the work directory of arXiv MCP. Do not point `WORK_DIR` to important folders. - Remote MCP connectors access MCP URL from the cloud. Local intranet addresses cannot be used for Claude remote connectors directly. - Remote MCP now prioritizes understanding through `http` transport. Do not continue using outdated SSE examples as the main path. - Paper abstracts, Chinese interpretation, and public account drafts are just auxiliary materials. Formal writing must return to the original text and citation verification results. - Do not submit API Keys, PDF originals, unauthorized full text, private notes, and `.env` to public GitHub repositories. ## 7. Keys and Security - `Claude Desktop login status`: This is your account status for using Claude client itself. - `SILICONFLOW_API_KEY`: This is the third-party model key used by `arXiv MCP Server` for parsing and generating Chinese content. - `PapersFlow` authentication: If the remote Connector requires login or authorization, complete it according to the webpage process. Do not fill in Anthropic or SiliconFlow keys by mistake. - Only put example values in public repositories, not real API Keys. - `WORK_DIR` is best to create a separate directory to avoid deletion and mis-transmission. ## 8. Repository Structure ```text . ├── README.md ├── claude_desktop_config.example.json ├── .env.example ├── .gitignore └── LICENSE ``` - `README.md`: The only main entry, installation, verification, and workflow are all here. - `claude_desktop_config.example.json`: Windows can directly modify and use local arXiv MCP example. - `.env.example`: Remind which variables need to be filled in, do not put real values. - `.gitignore`: Avoid submitting PDFs, output notes, and keys. - `LICENSE`: Clearly define the reuse boundary of the repository. ## References - [Claude Desktop Local MCP Server Description](https://support.claude.com/en/articles/10949351-getting-started-with-local-mcp-servers-on-claude-desktop) - [Claude Remote MCP Custom Connector Description](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) - [Claude Code MCP Official Documentation](https://docs.anthropic.com/en/docs/claude-code/mcp) - [arXiv MCP Server GitHub](https://github.com/yzfly/arxiv-mcp-server) - [arXiv MCP Server npm Package](https://www.npmjs.com/package/@langgpt/arxiv-mcp-server) - [PapersFlow MCP Server](https://mcpservers.org/servers/papersflow-ai/papersflow-mcp) - [PapersFlow Doxa Introduction](https://papersflow.ai/features/doxa)