Changelog

Changelog

0.9.0

• Use Pi’s native MCP configuration and agent-directory override, with the appropriate timeout units for native Pi and pi-mcp-adapter.
• Support CourtListener Token authentication and its default rolling quotas. Honor OpenAlex’s explicit shared daily-budget reset across processes, avoiding sibling-key retries when the provider’s client budget is exhausted.
• Register DeepSeek Harness through its native Cordis MCP plugin, preserving existing YAML and expressions. Add TRAE CLI 2.0 and MiMo Code through the shared TOML and OpenCode renderers, including their config-path overrides.
• The CPU reranker no longer holds gigabytes after a search: ONNX Runtime’s memory arena shrinks after every run (its own run option), and passages go in batches sized in tokens (4096 per run: 32 at 128 tokens, 10 at 384), because attention grows with batch times length squared. A pool of 400 read at 128 tokens and then its leaders at 384 peaked at 3.7 GB and now at about 0.45 GB, and it got faster (7 to 11 s to 6 to 9 s; one padded batch was also the slowest). The order moves less than the quantized model’s own noise from batch composition (Spearman 0.9987 and 9 of the top 10 against one batch of 200).
• Add a command line for research and setup: search, info, related, authors, read, download and graph call the same functions as the MCP tools; install registers Scholar in Pi, Claude Code, Claude Desktop, Codex, Cursor, VS Code, Gemini CLI, OpenCode, Antigravity, Windsurf, Kiro, Zed, Qwen Code, Kimi Code, Qoder CLI, CodeBuddy Code, iFlow CLI, MiniMax Code, Cline CLI, Kilo Code, Amp, GitHub Copilot CLI, Factory Droid, Junie CLI, Augment CLI, Warp and LM Studio with backups, atomic writes and dry-run diffs; uninstall, doctor (with --network probes), update and serve complete it. Running without arguments still serves MCP over stdio.
• Add scholar_mcp.api: the tools as Python functions returning dicts. Tools now return data and the MCP boundary renders YAML text plus structured content once.
• Accept every identifier Scholar returns, including OpenReview and Google Scholar ids, plus paper URLs and exact titles, in every tool. Reading and downloading reuse the merged record of a recent result.
• Keep ten-page reads whole in every client: read_paper declares a 120,000-character result size to Claude Code (which otherwise saves results over 50,000 characters to a file) and the installer raises Codex’s per-tool output limit (Codex had cut the middle out of a quarter of reads). A 100,000-character ceiling, cut at a page boundary with next_pages, guards unusually long ranges (SCHOLAR_READ_CHAR_BUDGET).
• The local ranking model reads the top 100 candidates a second time with their whole abstract, after a quick first pass over the pool.
• Fail fast on an unreachable remote reranker (10 s connect timeout) and rank locally for five minutes before retrying, instead of waiting out the full timeout on every pass.
• Write multi-line text as YAML literal blocks, which reads naturally and uses about 6% fewer tokens on paper text.
• Reuse private Google browser profiles and parse verified pages directly. Failed verification leaves no artificial retry freeze; optional host access can repair the request through its configured routes.
• Retain proven preprint/publication identities and prefer available versions of record. PDF upgrades keep separate files and preserve existing annotations. Temporary reading can reuse retained copies without saving a new file.
• Add native ERIC education search, CourtListener US case-law search and ClinicalTrials.gov registrations, protocols and linked papers, plus OpenEdition journals and books. Specialist records keep their own document types and identifiers.
• Ask Claude Code to load every tool upfront instead of behind tool search (about 2,800 tokens for all eight); "alwaysLoad": false in the server entry still defers them.
• Search every source in its own lane: several queries share one flow, each source answering them in order at its own pace (arXiv keeps its three-second spacing), and sources run side by side. Seed-title expansion and field discovery no longer wait for the slowest source between queries, and a source that misses the seed deadline still joins the final ranking.
• Citation and reference lookups and PDF resolution run in the same per-source lanes as searches, with the same deadline, so a stuck provider no longer holds paper_info or a read. Each provider’s published rate limit lives in its client, so metadata lookups keep the same spacing as searches.
• Search zbMATH Open, the reference index for mathematics, in every search. Its author summaries and expert reviews serve as abstracts for ranking.
• Add a global key file: scholar-mcp keys set NAME saves a source key in <data>/keys.env, which every client that starts Scholar reads.
• Add opt-in NASA ADS, Springer Nature, Web of Science Starter and Lens sources; each joins every search once its key is set and leaves installs without the key unchanged.
• Searches with intent="dataset" also search Zenodo’s datasets, ranked by best match, and searches written in Japanese also search CiNii Research; sources can now declare which searches they serve.
• Keep a bounded local record of tool calls (<data>/usage/, about 4 MB at most, no query text unless SCHOLAR_USAGE_LOG=full). doctor and scholar://status read it and name what changed for the worse: a source that started failing, went quiet, lost its abstracts or slowed down, a ranking model that fell back, reads that found no PDF, a tool that raised.
• A source that fails, for example out of credits or rate limited, is reported as failed rather than as a search that found no papers (Exa, arxiv.gg and OpenAlex citations and references used to answer with an empty list).
• scholar-mcp doctor --network probes every search, citation and reference source, the ranking model and the PDF chain live, and exits 1 on a problem. A daily workflow runs it and opens an issue when a probe fails.
• Hydrate expansion’s citation and reference results eight at a time instead of two; provider gates still bound the lookups (expansion 27 s to 19 s and 18 s to 13 s on the two profiled searches).
• Keep Europe PMC result formatting offline. A PMC dataset lookup ran for each search result and could hold a search for two minutes.
• Fetch publisher PDFs through the library proxy over plain HTTP first, with the installed browser’s user agent (old agents were bounced by IOP’s bot wall). Browser checks from Cloudflare, Radware and Elsevier are named in the failure reason alongside a library link to open in your browser. The proxy uses your exported cookies. Elsevier’s linkinghub hop maps to the article, and the Elsevier Article API is used when it reports full-text access.
• search_authors returns {authors: [...]}.

0.8.5

• Recover Google Scholar sessions automatically with the optional google extra, then reuse ordinary HTTP for searches and pagination. Verification is bounded, isolated in a temporary browser and coordinated across processes.
• Fall back to arXiv’s own HTTPS search when its Atom service is unavailable. Preserve original publication dates, authors and full abstracts.
• Handle bounded DBLP verification redirects and decode bibliography titles.
• Add scholar-mcp sources --check for live provider diagnostics without exposing credentials in error output.
• Correct Google Scholar metadata parsing around nonbreaking spaces.
• Include local reranking in MCPB builds and retain macOS Intel compatibility.
• Clarify installation profiles, reranker configuration and benchmark metrics.

Public tool names and schemas are unchanged.

Release artifacts, including the PyPI package, container images and desktop bundles, are listed on GitHub Releases.