dsh-ai4scholar

AI4Scholar for DeepSeek Harness

English | 中文

Academic literature tools for DeepSeek Harness (dsh), delivered as 43 native agent tools. Powered by ai4scholar.net.

Slash commands with natural-language requests

Choose a capability with a slash command, then describe the request normally. Unspecified criteria inherit the plugin settings.

Command	Purpose
/ai4s	Key status, balance, default sources/counts, and command index
/ai4s-search	Unified paper search; request multiple sources and deduplication in plain language
/ai4s-google	Google Scholar papers
/ai4s-semantic	Semantic Scholar search, bulk collection, and full-text snippets
/ai4s-snippets	Search matching passages in Semantic Scholar indexed full texts
/ai4s-pubmed	PubMed biomedical literature
/ai4s-arxiv	arXiv preprints
/ai4s-chinese	Chinese journals, theses, and conference literature
/ai4s-biorxiv, /ai4s-medrxiv	Browse preprints by category and time window
/ai4s-patents	Google Patents search
/ai4s-author	Author discovery, profiles, publications, and comparisons
/ai4s-paper	Details, title matching, batches, full text, PDF links, citations, references, recommendations, and citation formats
/ai4s-references	Outgoing references: papers this paper cites (Semantic Scholar)
/ai4s-citations	Incoming citations: papers that cite this paper
/ai4s-related	Related papers and single/multiple-seed recommendations
/ai4s-read	Read full text, including requested character slices
/ai4s-pdf	Obtain a PDF URL; this does not save a local file
/ai4s-export	Citation formats and BibTeX/export links for a known paper (Google Scholar)
/ai4s-cite	Add real references to prose
/ai4s-draw	Generate, edit, style, compose, iterate, critique, SVG, and vectorize scientific figures

There are 20 research commands plus /ai4s for account status, with /ai4scholar retained as an alias. Operation shortcuts disappear when their primary capability is disabled; lookup helpers alone do not keep them visible. Source and operation constraints are checked during tool execution. Snippets return excerpts, references are outgoing, citations are incoming, export formats an existing paper, and cite inserts references into prose.

/ai4s-google 10 solid-state battery papers published since 2022
/ai4s-search Search Google Scholar and Semantic Scholar for solid-state batteries and merge duplicates
/ai4s-author Use Semantic Scholar to find Geoffrey Hinton; confirm the author’s identity before showing the profile and representative papers
/ai4s-read Read the third paper above, focusing on its methodology
/ai4s-snippets Find 5 passages describing solid-state battery interface resistance experiments
/ai4s-references List the references cited by the first paper
/ai4s-citations Find papers citing the first paper
/ai4s-export Get the APA reference and BibTeX link for the first paper
/ai4s-patents Find CATL’s solid-state battery patent applications in China from the last five years

An empty command shows an example without searching. Research commands queue a normal model turn in the current conversation: your configured model interprets the request, briefly states the source and material criteria, then calls tools. Results remain available for follow-ups. Conflicting sources and ambiguous authors require clarification; source-specific commands enforce the selected source at execution and leave global defaults unchanged. /ai4scholar remains an alias for the status command.

On current Harness, paper, author, and patent result cards offer applicable actions. Buttons prepare an editable natural-language follow-up containing the exact identifiers; they preserve existing composer drafts. PDF/source links open directly. Older clients without the tool-view slot retain ordinary tool results.

Google Scholar results retain native paper, cited-by, and version identifiers. Citation-format lookup requires the native paper ID, not a DOI, URL, or author-page citationId. Author search defaults to Semantic Scholar; explicitly requesting Google Scholar selects its author tools. Google’s available article list is not guaranteed complete and supplies no h-index. bioRxiv/medRxiv browse by category and time, not free-text keywords. Patents stay separate from unified paper search.

Family	Tools	Cost
Paper search	search_papers — Semantic Scholar only by default; select PubMed, arXiv, Google Scholar, or Chinese literature via sources as needed. Multiple sources are merged by DOI / arXiv id / PMID / title, with multi-platform hits ranked first	per selected platform
Semantic Scholar (200M+ papers, all fields)	search_semantic, search_semantic_bulk, search_semantic_snippets (full-text excerpts), search_semantic_paper_match (title → paper), get_semantic_paper_detail, get_semantic_paper_batch, get_semantic_citations, get_semantic_references, get_semantic_paper_authors, search_semantic_authors, get_semantic_author_detail, get_semantic_author_batch, get_semantic_author_papers, get_semantic_recommendations, get_semantic_recommendations_for_paper, download_semantic, read_semantic_paper	credits
PubMed (biomedical & life sciences)	search_pubmed, get_pubmed_paper_detail, get_pubmed_paper_batch, get_pubmed_citations, get_pubmed_related	credits
Chinese literature (journals, theses, conferences; metadata and available abstracts)	search_chinese	credits
Google Scholar (papers, authors, reference formats)	search_google_scholar, search_google_scholar_profiles, get_google_scholar_author, get_google_scholar_citation	credits
Google Patents	search_google_patents	credits
arXiv	search_arxiv, download_arxiv, read_arxiv_paper	free
bioRxiv / medRxiv	search_biorxiv, search_medrxiv, download_biorxiv, download_medrxiv, read_biorxiv_paper, read_medrxiv_paper	free
Any DOI	download_by_doi, read_by_doi (open access anywhere; paywalled publishers on institutional networks)	free
Writing & figures	auto_cite (insert real citations + reference list + BibTeX), sci_draw (generate / edit / style / compose / SVG / vectorize scientific figures)	credits
Account	get_ai4scholar_credits, plus the /ai4scholar slash command	free

Start with one source. With the shipped settings, omitting sources (or passing []) queries only Semantic Scholar. If you choose other default platforms on the plugin page, the same call searches those platforms together. For biomedical or clinical topics, select sources: ["pubmed"] as appropriate. Use multiple sources, such as ["semantic-scholar", "pubmed"], when broader coverage is explicitly requested or one source is insufficient. Each billed API request adds to the charge; Google Scholar bills per page, while arXiv is free. If none of the configured default platforms is enabled, the default is the first enabled platform in this order: Semantic Scholar, PubMed, arXiv, Google Scholar, Chinese literature.

Every paper-list tool returns one normalized record shape — title, authors, year, venue, citation count, DOI/PMID/arXiv ids, URL, open-access PDF, abstract — so the model (and Code Mode programs) can chain calls across platforms without special cases. Results render as citation cards in the dsh Web UI; full-text tools return the PDF text in slices (offset / max_chars) so a 40-page paper never floods the context.

Credits are visible. Every billed tool card in the chat carries the charge in its title (Semantic Scholar: protein folding · 10 credits · 4,990 left), the model-facing result ends with AI4Scholar credits — this call: 10 · this session: 40 · remaining: 4,960 (from the API's X-Credits-* headers) and the model is asked to close a billed turn with a one-line credit note; /ai4scholar renders a balance card (available credits, breakdown, membership, key status, session spend) and get_ai4scholar_credits gives the model the same numbers. Free platforms (arXiv, bioRxiv/medRxiv, DOI) carry no credit line.

Chinese literature. Use search_chinese for Chinese keywords, document types (journal, thesis, conference), year filters, sorting, and pagination. It is also sources: ["chinese"] in search_papers, and stays unchecked in the shipped defaults. For a bilingual search, pass e.g. query: "bird migration", chinese_query: "鸟类迁徙", and sources: ["semantic-scholar", "chinese"]. Results share DOI and normalized-title de-duplication; short Chinese titles need the same year and first author, and conflicting DOIs stay separate. Translated titles without a shared identifier may remain separate. Chinese results provide metadata and available abstracts, not full-text downloads.

Screenshots

API key card (legacy Settings view)	/ai4scholar balance card

Install

Quick start (macOS / Linux; Windows PowerShell is the same):

# 0. Prerequisites: Node.js ≥ 22.19 (or ≥ 24) and pnpm (dsh manages plugins with pnpm)
node -v
npm i -g pnpm                     # skip if you already have it

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

# 2. This plugin (into the web profile; use --profile headless for the CLI runner)
dsh plugin --profile web add dsh-ai4scholar

# 3. Start
dsh web                           # prints http://127.0.0.1:3080

Then, once, in the browser:

1. Settings → Models — paste your DeepSeek API key (platform.deepseek.com).
2. Plugins → Installed → dsh-ai4scholar (or Settings → AI4Scholar) — paste your AI4Scholar API key (ai4scholar.net) and press Save. The card stores it in the dsh credentials store ($DSH_HOME/.credentials.yaml, mode 0600, the same place the Models page keeps provider keys), tests it at once (✓ Key works · 89,419 credits available), and it takes effect on the next tool call — no restart. Refresh re-checks any time. The saved key is masked by default, with Key configured on the right; click the eye to reveal or hide it. Hiding, leaving the page, or switching away from the window clears the revealed value.

On current DSH Desktop, click dsh-ai4scholar in the sidebar’s Plugins → Installed list: the key form is visible immediately on its detail page. Settings → Built-in plugins is a host inventory, so it no longer contains this form. Older Harness versions still offer the original Settings → Plugins → AI4Scholar card. Existing keys keep working; all entrances use the same credential reference.

Version 0.6.0 adds Ai4Scholar to the sidebar. Since 0.6.1, the sidebar uses a monochrome search icon that follows the theme and matches the other navigation icons. Its workbench searches directly without a chat model: choose a source, enter a query, and browse results beside the selected paper’s details. Settings in the workbench edits the same API key and defaults as the two existing entrances. Source-specific filters include result count, supported year ranges/date sorting, and the look-back window for bioRxiv/medRxiv category browsing. Full-text excerpts and patents have their own detail views. Results and recent queries stay in memory during the session; going back to an already loaded page does not make another request. Billed sources use API credits, including new pages; reported charges appear with the results. The workbench requires the authenticated plugin Fetch channel, validated on Harness 0.2.2-rc.1. Older hosts retain the existing tools and settings entrances.

Check for updates, beside Settings, checks the latest stable release on npm and shows both versions. A newer local test build is shown as ahead, never as a downgrade target. Release notes are linked; when a new version exists, Manage plugin opens the host’s plugin page. Checking does not install anything or use API credits.

Pick a workspace folder and ask away. On current Desktop, use Plugins → dsh-ai4scholar → Version to update a registry installation. Desktop manages its profile itself; the CLI does not modify it. To replace a local test package, uninstall it in Desktop and use Add plugin with dsh-ai4scholar@0.6.2, then enable it. Your API key stays in the shared credentials store. For the Web profile, use dsh plugin --profile web add dsh-ai4scholar@latest and restart dsh web.

Prefer files or CI? The plugin resolves the credential reference AI4SCHOLAR_API_KEY, so any of these also work (highest precedence first): export AI4SCHOLAR_API_KEY=… in the shell that launches dsh · a line in $DSH_HOME/.credentials.yaml · a line in the project or $DSH_HOME/.env. The card reports which layer is supplying the key and goes read-only when the environment does. Headless has no settings page, so use one of these there.

Publishing requests an npmmirror sync. Mirror availability is verified separately and may lag behind the official npm registry.

Verify the layer without booting: dsh --profile web --dump-config prints a # == dsh-ai4scholar section.

Then just ask:

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.

Configuration

On the plugin page

Open Plugins → AI4Scholar (or Settings → AI4Scholar; both edit the same values). Under the API key:

• Search: tick the platforms search_papers queries when the model names none (Semantic Scholar only by default), and set each platform's result count. Semantic Scholar, PubMed, and Chinese literature return up to 100 per search, billed once each; arXiv returns up to 100 for free. Google Scholar returns 1–10 per call (one credit); the model follows nextOffset with the same result count for more. All sources support pagination through their standalone tools; every additional billed request costs credits. The count is also the default for that platform's own search tool.
• Results: abstract length shown to the model (0–5000 characters, default 600), full-text slice per read_* call (2,000–200,000, default 60,000), and whether results carry the credit line.
• Command card: enable Show the command card after replies (off by default) to show the command menu below the latest completed answer. Click a command to fill the current conversation draft, add your request, then send. Existing drafts are preserved. Starting a new turn hides the previous menu; displaying it does not query AI4Scholar or the model. Requires the host's turn-tail extension and authenticated plugin Fetch channel; older hosts keep the manual /ai4s and /ai4scholar cards.
• Citations and figures: auto_cite reference style (IEEE by default) and whether to leave out preprints; sci_draw model (Flash 3.1 by default) and resolution (2K by default).

The model can still pass its own values for a single call. Save writes the changed fields into your profile's cordis.patch.yml and the plugin reloads with them. The next tool call and new conversations use the new values. Restore defaults removes them again. This page needs DeepSeek Harness 0.2.1 or later; on older versions it tells you to edit the file instead.

In the profile file

The bundle mounts one row (id: ai4scholar) with these defaults. Override from your profile's cordis.patch.yml (a patch replaces the whole config, so restate every key you keep). For DeepSeek Harness Desktop the file is ~/.dsh/profiles/desktop/cordis.patch.yml on macOS; for the CLI it is $DSH_HOME/profiles/<profile>/cordis.patch.yml:

- id: ai4scholar
  config:
    apiKeyEnv: AI4SCHOLAR_API_KEY   # credential reference; the key itself never lives in config
    baseUrl: https://ai4scholar.net
    # tool families
    semanticScholar: true
    pubmed: true
    googleScholar: true
    googlePatents: true              # search_google_patents
    chineseLiterature: true
    arxiv: true
    biorxiv: true                    # bioRxiv + medRxiv
    doi: true
    fullText: true                   # the read_* tools (PDF download + text extraction)
    autoCite: true
    sciDraw: true
    creditsTool: true                # get_ai4scholar_credits
    command: true                    # the /ai4scholar slash command
    balanceRoute: true               # GET /ai4scholar/balance for the settings card's key test
    showCredits: true                # credit line on billed results (plugin page)
    showCommandCard: false           # opt-in command menu after the latest reply (plugin page)
    promptGuidance: true             # short system-prompt section describing the tools
    promptOrder: 150
    # search_papers default platforms (plugin page); any of
    # semantic-scholar, pubmed, arxiv, google-scholar, chinese
    searchSources: [semantic-scholar]
    # per-platform result count (plugin page); a missing one uses defaultMaxResults
    resultsPerSource: {}             # e.g. { semantic-scholar: 100, google-scholar: 10 }
    citeStyle: ieee                  # auto_cite: ieee | apa | vancouver | nature | numbered (plugin page)
    citeExcludePreprints: false      # auto_cite (plugin page)
    drawModel: flash31               # sci_draw: flash | flash31 | pro | gptimage | gpt26flare | gpt26sunburst | banana21 | bananalite (plugin page)
    drawImageSize: 2K                # sci_draw: 1K | 2K | 4K (plugin page)
    # sizes
    defaultMaxResults: 10            # shared fallback count (configuration file)
    maxResultsCap: 100               # hard upper bound per call; each platform also caps
    abstractMaxChars: 600            # abstract characters shown per paper; 0 hides abstracts (plugin page)
    readMaxChars: 60000              # characters per full-text slice (plugin page)
    # timeouts / retry
    requestTimeoutMs: 30000          # per HTTP attempt
    pdfTimeoutMs: 120000             # per PDF download
    generationTimeoutMs: 300000      # auto_cite / sci_draw
    maxRetries: 3                    # attempts on 429 / network errors
    retryBackoffMs: 2000             # doubles per attempt
    toolTimeoutMs: 180000            # cooperative per-call budget enforced by dsh

The API key is resolved per call through ctx.credentials, so storing or rotating it takes effect on the next tool call without a restart. When the credentials service is absent (a custom composition), the plugin falls back to process.env[apiKeyEnv].

The settings card always edits AI4SCHOLAR_API_KEY (the browser half does not receive the row config); a deployment that overrides apiKeyEnv manages that reference through the environment or the credentials file instead. The session credit tally is process-local (it restarts with dsh); the balance itself always comes from the API.

How it fits dsh

• Two halves in one package: the Node half registers the tools; the browser half (exports["./client"], declared by dsh.client) contributes the settings card through native plugins.bundle.config / plugins.row.config slots and settings.section (plus settings.plugin.item on older hosts) writes keys through the credentials API. An explicit reveal uses POST /api/ai4scholar/key/reveal on the host’s authenticated, owner-only channel with Cache-Control: no-store; it is not a model tool and does not add the key to model context. Older hosts without this protected channel cannot reveal saved keys.
• Tools are registered on ctx.tools with defineTool: typed, schema-validated arguments; a canonical JSON value per call ({ source, query, total, papers[], truncated, nextOffset? }); a Markdown rendering for the model; and a web result card (structured sources) for the Web UI.
• A single tool:ai4scholar system-prompt section tells the model when to use which platform and to cite by DOI/link.
• Everything is an effect: unloading the plugin (or editing its config, which triggers HMR) withdraws every tool and the prompt section.
• No model-visible input is added outside the tool results, so sessions stay fully replayable from the log.

Development

For version numbers, test packages, and release checks, follow the release policy (Chinese). Routine development changes do not require a version bump.

pnpm install
pnpm test          # builds lib/ (Node half via tsc, browser bundle via tsdown), then runs:
                   #   unit tests, a smoke test against the real ToolRuntime, and a jsdom test that
                   #   loads the built client bundle under the dsh module-host contract and drives the card
pnpm typecheck

# try it in a dsh profile without publishing:
dsh plugin --profile web add /absolute/path/to/dsh-ai4scholar

dsh plugin add github:<you>/dsh-ai4scholar also works: the prepare script builds lib/ from source, and pnpm asks you to allow it once via allowBuilds in the profile's pnpm-workspace.yaml.

Notes

• download_by_doi / read_by_doi resolve through doi.org and known publisher PDF patterns (Elsevier, Springer/Nature, Wiley, T&F, MDPI, IEEE, ACM, ACS, RSC, PLOS, Frontiers, bioRxiv/medRxiv, arXiv). Paywalled articles only download when dsh runs on a network with institutional access; a refusal names the landing page so the user can fetch it by hand.
• Full-text extraction uses pdf-parse (pdf.js) in-process; scanned/image-only PDFs yield an explicit error rather than empty text.
• The same tool set ships for OpenClaw, Codex and Hermes.

License

MIT

/ai4s-draw / sci_draw also accepts quality (GPT: low/medium/high, plus xhigh/max on GPT Image 2.6) independently of image_size, and thinking_level (Gemini: banana21 minimal/medium/high; flash31/bananalite minimal/high). Omit quality for other models. Higher quality/resolution can cost more; the API validates and charges the selected combination. Existing calls that omit quality keep their previous behavior.