Librarium：并行开展多源 AI 研究

Multi-provider deep research CLI — fans out queries to multiple search/AI APIs in parallel

Librarium Librarium sends one research question to the search and AI sources you choose, then saves each answer, citation and cost in a Run you can resume. Website · Docs · Agent setup · Laravel package Requirements Node.js 22.12 or newer for the npm package. Homebrew and the standalone binaries include their own runtime. An API key for at least one source, for example EXA_API_KEY. Research calls the providers you select, and those calls can cost money. Install npm install -g librarium # or brew install jkudish/tap/librarium Standalone macOS and Linux binaries install with curl -fsSL https://raw.githubusercontent.com/jkudish/librarium/main/scripts/install.sh | sh. Check the result with librarium --version; it prints a 2.x version. First research librarium setup librarium setup --check librarium research "Compare managed PostgreSQL services" --only exa --dry-run setup is an interactive wizard for keys, source options and a default cost cap. Keys can be literal values or references such as env:EXA_API_KEY, and environment variables work without a config file. setup --check reads local configuration only and makes no provider calls. The dry run plans the request without calling a provider or saving a Run. Your Run id will differ: run_01M5ZQD37CXSGVA07PVQZ93MB5 planned (quick) Compare managed PostgreSQL services exa: queued, 0 citations, reported USD unknown Remove --dry-run to run it. Librarium saves the Run before the first provider call, then prints each source's status, citation count and reported cost. Runs go to ./agents/librarium under the current directory unless runs.path is set. librarium show --source exa prints that source's answer and citations from the latest Run. Deep research and resuming librarium research "Compare managed PostgreSQL services" --deep --only exa,openai --max-cost 1 --no-wait librarium show <run-id> --wait --json librarium summarize <run-id> --verify --deep submits durable research tasks. --no-wait returns after submission and requires --deep, because quick work runs inside the calling process. show resumes the saved task handles; it never submits the work again. Ctrl-C detaches and prints the show command to use later. summarize writes a separate, budgeted Run linked to the original by summary_of. Choosing sources Without --only, Librarium uses every research source you have set up, at quick depth. --only and --skip take comma-separated source, engine or saved group names. Naming a source without credentials is an error. There are 16 research sources: brave, claude, exa, firecrawl, gemini, google, grok, jina, kagi, openai, openrouter, parallel, perplexity, tavily, valyu and you. There are 5 engines: chatgpt, gemini-app, google-ai-mode, google-ai-overview and copilot. Engines record what a consumer AI product showed, collected through SearchAPI. They run only when named, and their results carry observation: "surface". Groups are lists you save yourself with groups add; none are built in: librarium groups add visibility chatgpt gemini-app google-ai-mode librarium research "How do AI products describe Acme?" --only visibility librarium groups remove visibility The sources command (librarium sources --json) shows what your installation can use, with options and price hints. The provider reference describes each source's quick and deep variants. Source keys lists the environment variable each source reads. Coding agents The mcp command (librarium mcp) runs a stdio MCP server with three tools: research starts a Run and returns a summary with statuses, citation counts and costs. get_run resumes a Run or pages through one source's answer and citations. list_sources lists sources, engines, saved groups and price hints. research and get_run wait 45 seconds by default and at most 50 (wait_seconds). An MCP max_cost can only lower the configured cap. The install guide has registration steps for each agent host. For Claude Code: claude mcp add --scope user librarium -- librarium mcp librarium install-skill installs the bundled SKILL.md into ~/.claude/skills/librarium. JavaScript npm install librarium import { createLibrarium } from 'librarium'; const librarium = createLibrarium(); const run = await librarium.research('Compare managed PostgreSQL services', { only: ['exa'], maxCost: '0.50', }); for (const result of run.results) console.log(result.source, result.status, result.citations.length, result.cost.reported); research resolves when the Run finishes. It prints one line per source; cost.reported is null when the provider did not report a charge. The root import reads ~/.config/librarium/config.json and stores Runs on disk. librarium/core runs in Workers and browsers with a memory store by default. See the library docs for get, sources, summarize, imports and custom RunStores. Other commands browse lists saved Runs. html and jsonl export a Run. usage totals reported spend. cleanup prunes finished Runs older than runs.keep_days. config reads and edits the config file. completions prints bash, zsh or fish completions. upgrade updates an npm, pnpm, yarn or Homebrew install, and install-skill installs the agent skill. The command reference has flags and examples. Limits A cost cap (--max-cost or max_cost) skips sources without a price estimate, then admits each source, in catalog order, whose estimate fits the remaining budget. It stops unstarted work, but calls already in flight can finish above the cap. Reported cost comes from the provider only. Unknown cost is null, never "0". A failed call may still be billed, and Librarium never substitutes another source after a failure. Provider answers and citations are untrusted evidence. Credentials are redacted before storage, but questions and cited pages can still contain private information. 2.0 migrates a 1.x config to version 3 after writing a backup. show cannot read 1.x run directories. Read the upgrade guide before upgrading. More documentation Configuration and source options Custom sources and RunStores Costs and limits