We treat lines as hypotheses and tool calls as measurements. Reasoning is not decoration of action — it is action, made visible at the speed of thought.
LV Agent is built on a Harness micro-kernel inspired by the Turing machine. It runs multi-turn LLM calls inside a tool loop with self-correction: simple questions need few turns, complex ones need many. A sketchbook is a research instrument — so is a terminal. A readable sketch beats a merely admirable render.
Multi-strategy reasoning — CoT / ReAct / Verify, with MCTS planning. Adaptive loop control: fewer turns for simple questions, more for complex ones.
Web search with multi-query fusion, full filesystem access, Python & Bash execution with timeout isolation, plus GitHub, PDF, weather, fetch.
Holographic memory: knowledge graph for persistent facts, vector experience memory across sessions, and learning skills extracted from conversations.
Traceable and replayable event-sourced execution. SQLite session persistence, dual token/time budget, checkpoint recovery, and hot-swappable modules.
Repeated tool hallucinations are corrected at parse time: MCP filesystem calls transparently map to native file ops, browser attempts fall back to web search. Fewer dead loops, faster responses.
Long conversations keep the latest tool results anchored at the tail of the dialogue, compressing history while preserving the most recent tool results — so "search → result → conclusion" never breaks.
Repeated observations fold into a single anchor point; tool failures auto-switch paths. Fewer wasted turns, verifiable conclusions arrive sooner on complex tasks.
Current builds are unsigned. On macOS, if blocked on first open, go to "System Settings → Privacy & Security" and allow; on Windows, extract and run LV Agent.exe. Bot/tool features require Python 3 and corresponding dependencies installed locally.
No installation required — open your browser and experience LV Agent's core capabilities: multi-turn reasoning, tool use, and real-time streaming output.
No dependencies to install — just open the page and start talking to LV Agent. Model and tool chain auto-configured.
Web search, file I/O, PDF generation, code execution — all tools available in the browser, with downloadable artifacts.
Thinking process, tool calls, and intermediate results stream in real time. Every reasoning step is visible and traceable.
Switch between sessions, review history, and download workspace artifacts with one click. Seamless across devices.
# Terminal edition
pip install lv-agent
# Or download desktop DMG
Supports macOS (arm64) and Windows (x64).
# Set API Key
export LV_API_KEY="sk-..."
First run auto-creates ~/.lv-agent/config.yaml.
lv
# Launch interactive terminal
Type a question to begin, /help for commands.
The web edition needs no setup — just visit lv-agent-web.onrender.com to start.
Full documentation and API reference at GitHub.
Default config uses DeepSeek-V4-Flash. Also supports OpenAI, Anthropic, Google Gemini, and other mainstream models — switch in config.yaml.
Pass it via the LV_API_KEY environment variable, or write it to ~/.lv-agent/config.yaml. The key is stored locally only and never uploaded to any server.
The terminal edition provides the full 19+ tool chain, local filesystem access, and desktop integration. The web edition runs in the browser, supporting core reasoning and tool use — ideal for quick trials and cross-device usage. See the comparison below.
No. LV Agent runs locally; conversations and tool calls go directly to your configured LLM API. Your code, files, and data never pass through LV Agent's servers.
macOS (Apple Silicon arm64) and Windows 10/11 (x64). Linux can use the terminal edition via pip install lv-agent. The web edition works in any modern browser.
Welcome to submit Issues or Pull Requests on GitHub. Please follow the project's code style and testing conventions.