LLMSX Explorer — The Dynamic Research Workbench

The Research Workbench

Explore any field you thought you knew

Most technical domains look familiar until you map their invisible frontiers, structural tradeoffs, and failure modes. LLMSX Explorer is an interactive, high-density research terminal and visualizer designed to turn static documentation into an accelerated learning loop.

Quick Install (Python 3.11+)

# Install with Textual TUI support
python -m pip install 'https://llms-explorer.com/downloads/llmsx/llmsx-0.2.6-py3-none-any.whl[tui]'

# Launch the terminal workbench (auto-clones the concept tree on first run)
llmsx explorer

Zero credentials required to browse. Zero telemetry. All notes and highlights persist locally on your machine.
New here? Read the complete user guide — it also ships in the app (press g, or run llmsx explorer --guide).

Dynamic and Focused Research

Traditional research starts with search engines or conversational models that reset state on every session. LLMSX Explorer operates on a durable, graph-aware concept tree. When you encounter a frontier concept or want to drill deeper into an existing node, dedicated research triggers are built right into the workbench:

Multi-Provider LLM & API Key Configuration

LLMSX Explorer is not locked to a single AI backend. You can dispatch research, skill runs, and braindump parsing across any supported engine:

• Anthropic Claude: claude -p with streaming JSON telemetry and tool execution.
• Google Gemini & AGY: gemini or agy CLI runner with stream-json output. Configurable via GEMINI_API_KEY or GOOGLE_API_KEY.
• OpenAI Codex: codex exec --json runner with model overrides. Configurable via OPENAI_API_KEY or CODEX_API_KEY.
• GitHub Copilot: gh copilot -- -p CLI integration using your GitHub token (GITHUB_TOKEN / COPILOT_API_KEY).
• Ollama (Local): Zero-cloud offline execution with any local model (e.g. llama3.2, deepseek-r1, qwen2.5-coder). Optional OLLAMA_API_KEY for remote instances.

Press , (comma) or click the toolbar provider badge (e.g. 🤖 claude) to open Settings. Select your active provider, configure model overrides, and enter API keys. Keys are stored locally at ~/.llmsx/config.json with strict 0600 file permissions, masked on screen, and never committed to git. Click Test Provider to verify CLI binary presence and credential readiness in one click.

Live Queue Viewer & Active Job Monitor

Press u or click 📋 Queue at any time to open the live Queue Viewer. The viewer splits into an active job status monitor and a pending queue table:

• Live Streaming Telemetry: View the currently executing prompt, session ID, model name, elapsed time, turn counter, and estimated dollar cost.
• Streaming Event Log: See tool executions (→ Read, → Bash), assistant thoughts, and status updates as they happen.
• Snapshot Protection: Before any research runner touches the repo, a git and tree snapshot is recorded. If a job crashes, times out, or is cancelled (x), the tree snapshot is atomically restored. Your repo is never left in a corrupted state.
• Queue Control: Launch any queued concept immediately (r), cancel running jobs (x), or open the raw job log (o).

Active Learning: Flashcards & Quizzes

Exploring a field is only the first half; retaining and testing that knowledge is the other. LLMSX Explorer extracts structured facts, claims, and definitions directly from researched concept packs to power active learning:

Freeform Capture: Braindumps & Daily Journals

Research is messy. You stumble across unexpected insights, transient debug logs, or raw ideas that don't fit neatly into an existing taxonomy yet.

• Braindump (W): Press W to open a capture window. Type or paste raw, unstructured thinking, code snippets, or conversational logs. The backend abstractor analyzes the text, extracts key concepts, and builds loadable .llms files.
• Concept Notes (n): Press n on any concept to open a markdown editor. Notes are stored in your home directory at ~/.llmsx/notes/<slug>.mdwith 0600 permissions. They stay private on your workstation and are never pushed to the shared repository.
• Daily Journal (J): Keep an append-only dated research journal tied to your active exploration sessions.

Facts, the llms Colour Scheme & the Scratchpad

The explorer opens every concept on its facts. The llms-family tabs (llms.txt,llms-full.txt, llms-small.txt, llms-facts.txt,llms-vocabulary.txt) show one numbered line per row, coloured by the llms format:

• Structure: # titles magenta, ## sections cyan, the > summary green, list markers yellow.
• Links: link names bold, URLs underlined blue, inline code orange.
• Provenance: [src: …] and [gen: …] tags and the — path:line (as of date) citation tail dimmed, dates green.
• Warnings: STALE:, Misbelief:, TODO, never, must in bold red.

The scratchpad is one global file, ~/.llmsx/scratchpad.md. Look up a thing, press y on each useful URL or fact, and every line is kept as- <line> — <concept> · <file>:<line> (as of <date>). On the Overview, Facts or Skill tab, y opens a filterable line picker instead. Duplicates are skipped. Press p (or 📝 Scratchpad) to reorder and edit, ctrl+sto save and ctrl+y to copy the whole list.

Unresearched concepts. Many concepts are known only by their place in the tree. Press U to mark every concept with no facts file further-research and add it to the research queue in one confirmed step; then filter with T and let🚀 Auto work through them.

Arrow keys. The outline, line views, lists and dropdowns keep their own↑ ↓. Everywhere else, ↓ moves focus to the widget directly below: from the filter box to the outline, from a button to the tab bar.

Highlights & Annotations Section

As you read through concept overviews, facts files, and skill references, you will find standout lines, critical architectural constants, or pithy summaries you want to keep.

Complete Hotkey Reference

Press ? at any time in the terminal workbench to bring up the interactive cheatsheet, or g for the complete user guide, which covers every screen, file and setting.

Dual Surface: Web & Terminal

LLMSX Explorer provides two complementary surfaces for engaging with the concept knowledge graph: