Learn Crucible

Crucible needs a language model. The easiest free options are LM Studio (lmstudio.ai — point-and-click, download a model from inside the app) or Ollama (ollama.ai, then `ollama pull gemma3:4b` in a terminal). Already have an OpenAI or Anthropic API key? Skip to Step 3.

Double-click the Crucible shortcut on your desktop. A native window opens, so there is no browser tab and no terminal to keep open.

Nothing to do — v0.3.5 opens straight to your vault with every feature unlocked. The sign-in and account panels are still visible in Settings but greyed out, kept only in case accounts return for cross-device sync. Older builds asked you to click "Continue without account"; that screen is gone.

Go to Settings and pick the Local provider — it covers Ollama, LM Studio, llama.cpp and vLLM. Crucible scans for a running server on launch and fills in the details, so usually there is nothing to type. To set it by hand: Ollama uses base URL http://localhost:11434, LM Studio uses http://localhost:1234/v1 (the /v1 suffix is what tells Crucible to speak the OpenAI format). For cloud APIs: pick OpenAI or Anthropic and paste your key. Click Test Connection to verify.

Go to the Interests page. Type what you want to research in plain English (e.g. "Global energy markets, AI regulation, semiconductor supply chains"). Click Generate and the LLM converts it into structured research topics.

Go to the Dashboard and click Run Now. With the default limits and a quick model, your first batch of articles is graded, summarized and filed in roughly 5 to 15 minutes. Deep runs are a different animal: several hundred URLs and documents, on a local model producing 30 to 40 tokens per second, can take two to eight hours. Both are normal. Start small, then raise the limits in Settings once you know what your machine gets through in an hour.

Your home base. Shows total documents, topic breakdown with counts, average quality score, source count, token usage per cycle, recent documents with clickable titles, and a live activity feed.

Browse collected articles by topic. Each article has a title, source URL, quality score (1–10), topic, LLM-written summary, tags, and full cleaned text. Delete documents that don’t belong.

Find any article by keyword. Full-text search across titles, summaries, and content. Results are instant thanks to the SQLite FTS5 index.

Every website Crucible has scraped, with trust tiers (high/medium/low), doc counts, quality averages, and success rates. Promote good domains, block bad ones, unblock when needed.

See how your vault is organized. Each topic shows its document count. Edit the taxonomy directly to add, remove, or restructure categories.

Your structured research goals. Add, remove, or regenerate topics any time. Changes take effect on the next cycle.

Visual map of how documents connect: causal links, contradictions, thematic ties, temporal chains. Explore clusters, browse narrative chains, and view edge statistics. Use Reconnect after cycles to discover new edges.

Log of every collection cycle: documents fetched/kept/rejected, token usage, estimated cost, and errors. Click a run for step-by-step details.

Configure LLM provider and model, temperature, context window, per-step model overrides, scrape delays, blocked domains, and research templates. Test Connection verifies your LLM before you run.

Pre-built profiles for Geopolitical Intelligence, Business Intelligence, Academic Research, and Sports Analytics. Load one to populate your goals, topics, and feeds instantly.

Start fresh: deletes all documents, index, graph, and activity, but keeps your configuration, interests, and templates.

Your vault is plain markdown files. Open the folder in Obsidian for advanced note-taking alongside collected research. Wiki-links between related articles work natively.

For fully private inference, run a model on your own machine. LM Studio (http://localhost:1234/v1) and Ollama (http://localhost:11434) are the easiest; llama.cpp, vLLM and Jan work too. Choose the Local provider in settings — Crucible detects a running server automatically and picks the right protocol from the base URL. A GPU with 4+ GB VRAM is recommended, though small models run on CPU. We recommend Bonsai 27B, a low-bit build of Qwen3.6-27B: the 1-bit GGUF is a 3.8 GB file and the ternary 2-bit build is 7.2 GB, so a 27B model fits in roughly the memory a small one usually asks for, and it reads and grades documents far better than a 4B. Search for Bonsai-27B in LM Studio. The model is what does the judging, so it sets the ceiling on how good your vault gets.

Add your OpenAI API key in settings. Crucible uses GPT models for document evaluation, summarization, and knowledge graph construction. Per-step model overrides let you use cheaper models for bulk tasks.

Add your Anthropic API key in settings. Claude models are used for document analysis and relationship extraction. Recommended for high-quality knowledge graph edges.

Included. Privacy-respecting web search with no API key required. Good for general research topics.

Included. Requires your own Brave Search API key. Provides independent web index results that complement DuckDuckGo.

Included. Meta-search engine that queries dozens of search engines simultaneously through a privacy-respecting proxy. Self-hosted or use a public instance.

Included. Direct integration with the arXiv API for discovering academic papers. Supports citation extraction and structured metadata.

Every document is stored as a markdown file with YAML frontmatter containing metadata: source URL, quality score, categories, timestamps, and relationship data. Fully compatible with Obsidian.

Documents are connected via typed edges: causal, temporal, thematic, and contradictory. Multi-hop traversal lets you discover chains of reasoning across your entire vault.

Included. Export the entire vault as a ZIP archive for backup, migration, or offline access.

Included. Trigger a single research cycle on demand. The AI director plans queries, executes searches, scrapes pages, evaluates relevance, and stores results.

Included. Set daily, weekly, or custom cron intervals for automatic collection cycles. Crucible runs in the background on your schedule.

Included. Always-on collection mode that continuously monitors your topics of interest, discovering and ingesting new sources as they appear.

Monitor competitors, track industry trends, and aggregate market data automatically. Crucible continuously scrapes press releases, filings, news, and social signals into a structured vault you can query and analyze.

Pull papers from arXiv, government databases (FRED, EIA), and institutional repositories. The knowledge graph surfaces citation chains and thematic connections across hundreds of papers without manual curation.

Build information edges by aggregating diverse signals from news, data feeds, and expert analysis. Crucible's structured vault makes it easy to track evolving narratives and export curated datasets for modeling.

Use Crucible as the collection layer upstream of your analysis stack. Export structured data to Jupyter, Tableau, or custom pipelines. The YAML frontmatter and knowledge graph metadata make datasets immediately queryable.

Hobbyists, writers, and lifelong learners use Crucible to build deep knowledge bases on any topic. The Obsidian-compatible vault integrates with your existing notes, and scheduled collection keeps your vault growing while you sleep.

Make sure your LLM is running. For Ollama: run `ollama serve` in a terminal first. For LM Studio: start the local server from its Developer tab and check the base URL ends in /v1. For cloud APIs: verify your API key is correct and has credits.

Make your interests more specific. Vague topics produce poor search queries. Try narrowing from 'technology' to 'semiconductor supply chain disruptions in 2025'.

Try a smaller model (gemma3:4b instead of llama3:70b), increase timeout in Settings, or reduce documents per cycle.

Block bad source domains, lower temperature to 0.2 for more focused evaluation, and promote domains you trust.

You need at least 5 documents before the graph can find connections. Run Reconnect after a few cycles to discover new edges.