Foreword¶

When conducting literature research in DeepSeek Harness (dsh), the common practice is to manually open Semantic Scholar, PubMed, or arXiv, copy the titles and abstracts back into the conversation, and then have the model summarize them. When searching across platforms, you also need to manually merge duplicate entries and align DOIs and citation formats. The steps are scattered, and the context window can easily be filled up.

Below, we introduce dsh-ai4scholar (maintained by literaf, GitHub repository literaf/dsh-ai4scholar), which integrates academic search capabilities as native agent tools in dsh. After a single installation, the model can directly call the interfaces of Semantic Scholar, PubMed, Google Scholar, arXiv, bioRxiv/medRxiv, DOI resolution, and more. It also supports full-text reading, automatic citation generation, and scientific plotting. The backend is provided by ai4scholar.net and shares the same capabilities as the AI4Scholar plugins on OpenClaw, Codex, and Hermes.

What This Is¶

dsh-ai4scholar is a web-connected plugin for DeepSeek Harness (npm package name: dsh-ai4scholar, current version v0.3.3, MIT license). It integrates into dsh as 38 native tools, covering literature search, details and citation relationships, PDF full-text reading, automatic citation (auto_cite), and scientific plotting (sci_draw).

All tools that return lists of papers use a unified record structure (title, authors, year, venue, citation count, DOI/PMID/arXiv ID, URL, open-access PDF, abstract), making it easy for the model to chain calls across different platforms. Full-text tools return PDF text in segments based on offset / max_chars to prevent long texts from overwhelming the context. Results from billing tools are displayed as citation cards in the Web UI, along with credits consumption information.

Core Features¶

Unified Search¶

search_papers can query across Semantic Scholar and PubMed (with optional arXiv and Google Scholar) in a single call, deduplicates results by DOI / arXiv ID / PMID / title, and prioritizes entries that match across multiple platforms.

Platform-Specific Tool Families¶

Semantic Scholar tools include search_semantic, search_semantic_bulk, search_semantic_snippets, get_semantic_citations, get_semantic_references, read_semantic_paper, and more. PubMed, arXiv, and bioRxiv/medRxiv each have corresponding search / get / download / read tools. The full list is available in the GitHub README.

Transparent Credits¶

Billing tools display the credits consumed and remaining in the chat card title (e.g., Semantic Scholar: protein folding · 10 credits · 4,990 left). Model-side results include a suffix: AI4Scholar credits — this call: … · this session: … · remaining: …. The /ai4scholar command renders a balance card; get_ai4scholar_credits returns the same figures to the model. Free tools related to arXiv, bioRxiv/medRxiv, and DOI do not display credits information.

Installation and Setup¶

Requirements: Node.js ≥ 22.19 (or ≥ 24), pnpm, and dsh installed globally.

# 0. Check Node version
node -v
npm i -g pnpm                     # Skip if already installed

# 1. Install DeepSeek Harness
npm i -g @deepseek-ai/dsh
dsh --version

# 2. Install the plugin (Web profile; use --profile headless for headless CLI)
dsh plugin --profile web add dsh-ai4scholar

# 3. Start the Web UI
dsh web                           # Default: http://127.0.0.1:3080

After starting, complete the one-time configuration in the browser:

1. Settings → Models — Enter your DeepSeek API Key (platform.deepseek.com).
2. Settings → Plugins → AI4Scholar — Enter your AI4Scholar API Key (ai4scholar.net) and save. The key is stored in dsh’s credential store ($DSH_HOME/.credentials.yaml, permissions 0600). Saving triggers a validation via the plugin route GET /ai4scholar/balance, and the key takes effect on the next tool call without requiring a restart.

When there is no Web settings page (e.g., headless), you can inject the key via environment variables or files, with the following priority (high to low): export AI4SCHOLAR_API_KEY=… → $DSH_HOME/.credentials.yaml → .env in the project or $DSH_HOME. Use dsh --profile web --dump-config to view the # == dsh-ai4scholar configuration section.

To upgrade the plugin:

dsh plugin --profile web add dsh-ai4scholar@latest

Then restart dsh web. After release, the package is synced to npmmirror, and the new version is typically available within about a minute using the default registry in China.

Typical Usage¶

Select a workspace folder, then ask questions in natural language. Examples from the README:

Find recent papers on CRISPR base editing for sickle cell disease and compare their delivery methods.

Search PubMed for GLP-1 receptor agonists and cardiovascular outcomes since 2022, sorted by date.

Which paper is “Attention Is All You Need”? Give me its DOI and citation count.

The model will automatically choose tools like search_papers, search_pubmed, search_semantic based on the task. For full text, it will call tools like read_arxiv_paper or read_by_doi, which are subject to the readMaxChars limit (default 60,000 characters). During writing, you can have the model call auto_cite to generate a citation list with BibTeX.

The slash command /ai4scholar can be used at any time to check the current account balance, membership status, and session consumption.

Configuration Options¶

The plugin is mounted in the profile as id: ai4scholar. Default values can be overridden via cordis.patch.yml (the patch replaces the entire config, so all keys to be retained must be specified):

- id: ai4scholar
  config:
    apiKeyEnv: AI4SCHOLAR_API_KEY
    baseUrl: https://ai4scholar.net
    semanticScholar: true
    pubmed: true
    googleScholar: true
    arxiv: true
    biorxiv: true
    doi: true
    fullText: true
    autoCite: true
    sciDraw: true
    creditsTool: true
    command: true
    balanceRoute: true
    showCredits: true
    defaultMaxResults: 10
    maxResultsCap: 50
    abstractMaxChars: 600
    readMaxChars: 60000

You can disable specific platforms or feature families as needed (e.g., keep only arXiv and DOI) to reduce the number of tools visible to the model.

Use Cases and Notes¶

Who is this for: Researchers and developers conducting systematic literature reviews, cross-database comparisons, reading open-access full texts, organizing citations, or generating scientific figures in dsh. For Google Scholar citation counts or Semantic Scholar citation graphs, you will need an AI4Scholar API Key and should monitor credits consumption; for purely arXiv / preprint / open-access DOI scenarios, you can use the free tool families.

Running permissions: The plugin runs with the permissions of the current dsh process and can access the network and local workspace. Before installation, it is recommended to read the source code and the MIT license to confirm that the tool call scope aligns with your security policies.

Ecosystem note: DeepSeek Harness adopts an “everything is a plugin” architecture. Community directories like SkillHub are maintained by third parties and have no official affiliation with DeepSeek or High-Flyer. This plugin depends on the ai4scholar.net backend, with some capabilities billed per platform. Free platforms do not require API quota for search and reading tools.

Conclusion¶

dsh-ai4scholar consolidates scattered academic search, full-text reading, citation generation, and plotting capabilities from multiple sites into 38 orchestrable tools within dsh. It unifies data structures and transparently displays credits in the UI layer. If you are already using DeepSeek Harness for research assistance, you can follow the steps above to install and configure the API key, then use it directly in workspace conversations.

• Community directory: skillhub.cn/plugins/literaf/dsh-ai4scholar
• Source code and documentation: github.com/literaf/dsh-ai4scholar