research-hub

research-hub

research-hub turns Zotero, Obsidian, and NotebookLM into an AI-operable research workspace. It works best with any two of the three tools, and unlocks the full loop when all three are connected.

Runtime contract (do this first)

This is a runtime-backed skill. The SKILL.md instructions teach the AI workflow, but the actual search, Zotero writes, Obsidian vault updates, NotebookLM bundle/upload/download, dashboard, and MCP tools come from the research-hub Python CLI.

Before running any workflow command from this skill, check the runtime:

research-hub describe --json
research-hub doctor

If research-hub is not found, the host has loaded the skill instructions but the executable runtime is missing. This can happen after a marketplace/manual skill install without the Python package. Stop and tell them:

This skill needs the research-hub CLI. Please run:

pip install research-hub-pipeline
research-hub setup --persona researcher   # or analyst | humanities | internal

Then re-run your request. If you only need to compare papers, sharpen a research question, or build a project / paper memory file (no automated search, no NotebookLM upload), a lightweight prompt-only skill may be enough; this workflow needs the CLI for tool execution.

If doctor runs but reports missing Zotero credentials, missing NotebookLM auth, or no supported LLM CLI, adapt instead of pretending the full path is ready:

• Missing Zotero: use import-folder, sample/dashboard mode, or ask the user to set ZOTERO_API_KEY / ZOTERO_LIBRARY_ID.
• Missing NotebookLM auth: run the first search with --no-nlm, then ask the user to complete research-hub notebooklm login --auto-detect.
• Missing LLM CLI for the relevance judge: either install/configure a supported CLI (claude, codex, gemini, opencode, aichat, cursor, or a custom adapter) or add --no-fit-check and clearly state that relevance filtering was skipped.

Do not invent or simulate research-hub output if the CLI is missing.

Default language policy: answer the user in their language. Generate durable research notes, metadata, and citations in English unless the user explicitly asks for another language.

Agent preflight protocol

For AI hosts such as Claude Code, Codex, Gemini CLI, Cursor, OpenClaw, Hermes, or generic API clients, use this sequence unless the user gives a narrower command:

1. Capability check:

   research-hub describe --json
   research-hub doctor
   

2. First safe literature run, without NotebookLM:

   research-hub auto "TOPIC" --max-papers 3 --no-nlm
   

3. If the run stops before search because no relevance judge is on PATH, choose one of these explicit paths:

   research-hub auto "TOPIC" --max-papers 3 --no-nlm --no-fit-check
   research-hub auto "TOPIC" --max-papers 3 --no-nlm --llm-cli codex
   research-hub auto "TOPIC" --max-papers 3 --no-nlm --llm-cli gemini
   

4. Add NotebookLM only after the local Zotero/Obsidian path works:

   research-hub notebooklm login --auto-detect
   research-hub notebooklm bundle --cluster <slug>
   research-hub notebooklm upload --cluster <slug>
   research-hub notebooklm generate --cluster <slug> --type brief
   research-hub notebooklm download --cluster <slug>
   

5. For machine-readable automation, prefer commands with --json when available, or use the MCP/REST surfaces exposed by research-hub serve.

This protocol is intentionally staged: verify runtime first, ingest a small cluster second, then add browser-dependent NotebookLM work last.

Pick The Right Entry Point

User setup	Recommended path
Zotero + Obsidian + NotebookLM	research-hub auto "topic"
Zotero + Obsidian only	research-hub auto "topic" --no-nlm, zotero backfill, Obsidian dashboard output
Obsidian + NotebookLM only	research-hub import-folder <folder> --cluster <slug>, then NotebookLM bundle/upload
Zotero + NotebookLM only	Zotero-backed search and NotebookLM operations
No accounts yet	research-hub dashboard --sample

Setup Commands

pip install research-hub-pipeline[playwright,secrets]
research-hub setup
research-hub doctor

For local files without Zotero:

pip install research-hub-pipeline[import,secrets]
research-hub setup --persona analyst
research-hub import-folder ./papers --cluster my-local-review

Core Workflows

Preview

research-hub dashboard --sample

Research Topic

research-hub plan "TOPIC"
research-hub auto "TOPIC" --no-nlm
research-hub serve --dashboard

Use --no-nlm for first-run smoke tests or when NotebookLM browser automation is not configured.

Discover (search + AI fit-check)

Two-phase interactive flow when you want a human / AI in the loop on which papers actually belong in a cluster. Replaces the auto one-shot ingest when topic boundaries are fuzzy.

research-hub discover new --cluster project-topic --query "agent-based modeling flood adaptation"
# → emits search results + a fit-check scoring prompt; stashes state

# Run the fit-check prompt through your AI of choice, paste the scores back
research-hub discover continue --cluster project-topic --scores scores.json
# → applies scores, emits papers_input.json for ingest

research-hub fit-check {emit|apply|audit|drift} exposes the underlying gates separately when you want to re-score an existing cluster or audit drift over time. discover variants emits a query-variation prompt to widen recall before fit-check narrows it.

Local Source Folder

research-hub import-folder ./sources --cluster project-topic
research-hub serve --dashboard
research-hub crystal emit --cluster project-topic

NotebookLM

research-hub notebooklm login --auto-detect
research-hub notebooklm bundle --cluster project-topic
research-hub notebooklm upload --cluster project-topic
research-hub notebooklm generate --cluster project-topic --type brief
research-hub notebooklm download --cluster project-topic

Synthesize cluster pages

Generate or refresh per-cluster synthesis pages in the Obsidian vault (uses cluster memory + paper summaries to produce a navigable overview note).

research-hub synthesize --cluster project-topic
research-hub synthesize --cluster project-topic --graph-colors  # also paint the graph view

Run synthesize after paper-summarize has filled the per-paper notes; the synthesis page reads from those.

Cluster memory

Maintain a structured memory registry per cluster — durable notes the AI can reload across sessions without re-reading every paper.

research-hub memory list --cluster project-topic
research-hub memory read --cluster project-topic
research-hub memory emit --cluster project-topic   # AI extraction prompt
research-hub memory apply --cluster project-topic --payload memory.json

Use memory emit/apply to refresh the registry after a major round of new papers; use read from another session to reload context cheaply.

Maintenance

research-hub doctor --autofix
research-hub tidy
research-hub clusters rebind --emit
research-hub cleanup --all

MCP Integration

For MCP hosts:

{ "mcpServers": { "research-hub": { "command": "research-hub", "args": ["serve"] } } }

Install host-specific files for hosts with known default skill directories:

research-hub install --platform claude-code
research-hub install --platform cursor
research-hub install --platform codex
research-hub install --platform gemini

For Hermes, OpenClaw, or other hosts with SKILL.md/rules support, copy the relevant skills/<name>/ directories manually or inline this file into the host's instructions. Use MCP/REST for tool calls when no installer target exists.

Guardrails

• Do not assume installing ai-research-skills or copying SKILL.md installed the research-hub CLI; check the runtime explicitly.
• Always run research-hub doctor when setup state is uncertain.
• Do not invent DOIs, citations, or paper metadata; use search/enrich/verify commands.
• Do not delete clusters without reviewing cascade impact.
• Treat the vault as user-owned local data; avoid overwriting notes unless asked.
• Prefer import-folder for non-academic or internal documents.
• Prefer Zotero-backed workflows for DOI/arXiv-heavy academic literature.