Opencode Research Mcp

research-mcp

A lean research MCP that bundles academic search, citation graph traversal, OA full-text download, and browser-based institutional download into 5 tools. Built on OpenAlex, Semantic Scholar, CrossRef, and 6 other academic indexes.

What it does: You ask a research question, it searches 8+ academic databases in parallel, removes duplicates, ranks by semantic + keyword relevance, and returns the best papers. Citation walk follows the graph forward and backward via OpenAlex. read_paper fetches OA full text. browser_download uses a real Playwright browser session for institutional SSO paywalls.

Why it's lean: 5 tools are still tiny compared with the ~12,000-token surface of the underlying MCPs. Per the DADL framework, each tool adds context pressure, so we keep the surface focused: search, citations, OA read, browser paywall read, and ping.

5 Tools

| # | Tool | Purpose | |---|------|---------| | 1 | search_literature | 8+ sources, dedup, semantic+keyword ranking, mode + field + debug params | | 2 | walk_citations | Multi-hop citation graph (forward/backward/both) via OpenAlex | | 3 | read_paper | OA-first full-text download with auto-detect + optional Sci-Hub fallback | | 4 | browser_download | Playwright browser download through institutional SSO/EZproxy, with PDF title verification | | 5 | ping | Cheap health check; use before expensive search/download calls |

Search Modes

search_literature has a mode parameter that changes ranking and filtering without changing the response shape (zero context bloat):

| Mode | Filter | Ranking | |------|--------|---------| | "comprehensive" (default) | — | Source tier → source hits → relevance → abstract → citations → year | | "seminal" | ≥10 citations | Citations desc, oldest first | | "recent" | Last 2 years | Citations desc, newest first | | "survey" | Review/survey/meta-analysis only | Relevance → survey flag → citations |

Field-Aware Source Tiering

search_literature accepts field="auto" (default) which detects the query's field and applies field-aware source weighting. Sources are tiered:

| Tier | Sources | When Counted | |------|---------|--------------| | 3 (top) | semantic, scopus, openalex, crossref, springer | All fields | | 2 (mid) | arxiv, unpaywall, openaire, doaj | All fields | | 1 (field-specific) | europepmc | Only for medical and bio fields |

In CS/AI queries, Europe PMC papers are demoted to tier 0 (biomedical venue, irrelevant). In medical queries, Europe PMC papers get full tier-1 credit. The field param accepts "cs", "bio", "medical", "social", "general", or "auto".

The safe filter (_should_drop_low_quality) drops papers with no abstract AND /download_ledger.jsonl`.

• Reuse existing PDFs: if the ledger has a verified PDF for the DOI/title, the tool returns it without opening the browser again.
• Retry backoff: recent failures block immediate retry unless force=True. Title mismatches are blocked for 7 days; login/timeouts use shorter cooldowns.
• Verification-first: success requires extracted PDF text to match requested title keywords. Wrong PDFs are deleted automatically.

Useful params:

browser_download(..., reuse_existing=True, force=False, human_delay=True)

For tests on public PDFs only, set human_delay=False. For real institutional downloads, keep human_delay=True.

Sources

| Source | Type | Key Required? | |--------|------|---------------| | arXiv | Preprints | No | | Semantic Scholar | Academic search | Recommended | | OpenAlex | 270M+ publications | No | | CrossRef | DOI resolution | No | | Unpaywall | OA PDF resolver | Email recommended | | OpenAIRE | EU open science | No | | Europe PMC | Biomedical | No | | DOAJ | Open access journals | No | | Scopus | 26K+ journals | Elsevier API key | | Springer Nature | 29M+ papers | Springer API key |

Design Rationale

This design is grounded in the MCP tool selection literature:

• Wang et al. (2026) — Tool catalog size inversely correlates with selection accuracy. MCPs with >40 tools see -260% selection quality vs <15 tools.
• Dunkel (2026) — DADL framework: context window grows linearly with tool catalog size. Each tool adds ~1.5% context pressure.
• Hou et al. (2026) — MCP security analysis: bloated tool surfaces create 16 attack vectors through dangling or misdescribed tools.
• Gan & Sun (2025) — RAG-MCP: tool routing quality degrades by 12% per 10 tools. Bundled servers with <10 tools achieve 89% routing accuracy.

Agent Config

The agents/researcher.md file configures the researcher subagent for use with this MCP. Key settings:

• task: deny — Prevents the researcher from spawning sub-sub-agents (avoids runaway cost spirals)
• Capped work scope — Max 2 search + 2 walk + 1 read + 1 browser_download call per request
• Stale tool references removed — No research_paper_lookup (deleted in Round 1)

Install: copy agents/researcher.md to ~/.config/opencode/agents/researcher.md.

License

MIT

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: chessy795
• Source: chessy795/opencode-research-mcp
• License: MIT

Install and usage instructions live in the source repository linked above.