Librarium -- Multi-Provider Deep Research

Run research queries across 10 search and deep-research APIs in parallel, collect results, deduplicate sources, and produce structured output.

Prerequisites

• librarium CLI installed (npm install -g librarium)
• API keys configured (librarium init --auto)
• Binary at: librarium (or npx librarium)

7-Phase Research Workflow

Phase 1: Query Analysis

Analyze the user's research question. Determine:

• Is this a technical, business, or general knowledge query?
• Which provider group is best suited? (quick for fast answers, deep for thorough research, comprehensive for important decisions, all for maximum grounded coverage, llm for an ungrounded baseline/contrast with no citations)
• What execution mode? (sync for quick queries, mixed for deep research)

Phase 2: Provider Selection

Select providers based on query type:

• Technical queries: Use comprehensive group (deep research + AI-grounded)
• Quick facts: Use quick group (AI-grounded only, fast)
• Competitive research: Use all group (maximum coverage)
• Specific provider: Use --providers flag (accepts canonical IDs or display names, e.g. -p "Exa Search,brave-search")
• Competitive research: Use all group (maximum grounded coverage)
• Ungrounded baseline / contrast: Use llm group (Claude, OpenAI, Gemini, OpenRouter -- direct model answers, no citations)
• Specific provider: Use --providers flag

Phase 3: Dispatch

Run the query:

librarium run "your query here" --group <group> [--mode mixed]

Phase 4: Monitor Async Tasks

If deep-research providers were used in async mode:

librarium status --wait

Phase 5: Retrieve Results

Once complete, async results can be retrieved:

librarium status --retrieve

Phase 6: Analyze Output

Read the output files:

1. summary.md -- Overall research summary with statistics
2. sources.json -- Deduplicated citations ranked by frequency
3. Individual {provider}.md files for detailed per-provider results
4. run.json -- Machine-readable manifest

Phase 7: Synthesize

Combine findings from multiple providers into a coherent answer. Cross-reference sources that appear across multiple providers (higher citation count = higher confidence).

Key Commands

MCP Server

Instead of shelling out to the CLI, agents can drive librarium over the Model Context Protocol with librarium mcp (stdio transport). Register it once with claude mcp add librarium -- librarium mcp, then call the tools: research, get_results, check_async, list_providers, list_groups. The research tool runs the same silent file-writing pipeline as librarium run and returns a compact structured result; fetch full provider markdown with get_results.

Provider Tiers

Output Structure

./agents/librarium/{timestamp}-{slug}/
  prompt.md, run.json, summary.md, sources.json
  {provider}.md, {provider}.meta.json
  async-tasks.json (if applicable)