LV LV Agent v1.0.0 · macOS · Cleveris Research
§ 01

We treat the line as a hypothesis and the tool call as a measurement. Reasoning is not decoration for action — it is action, made visible at the speed of thought.

LV Agent is built around a Harness microkernel inspired by the Turing machine. It runs multi-round LLM calls inside a tool loop with self-correction: simple questions take few rounds, hard questions take many. The sketchbook is a research instrument — and so is the terminal. A sketch that can be read is worth more than a render that can only be admired.

Sketch study — hands approaching
§ 02 Four plates of the same instrument
FIG. II

Reasoning & planning

Multi-strategy reasoning — CoT / ReAct / Verification, with MCTS planning. Adaptive loop control: few rounds for easy asks, many for hard ones.

  • SEQUENTIAL · PARALLEL · HIERARCHICAL · ADAPTIVE
  • Self-correction with quality evaluation
  • Parameter auto-tuning per task
FIG. III

Native tool chain

Web search with multi-query fusion, full file system access, Python & Bash execution with timeout isolation, plus GitHub, PDF, weather, fetch.

  • Read / write / grep / glob, cross-folder
  • Pipes, redirect, && / ||, env passthrough
  • Telegram bot integration (standalone)
FIG. IV

Memory & context

Holographic memory: a knowledge graph for durable facts, vector experience memory across sessions, and learned skills extracted from conversation.

  • Entity-relation long-term store
  • Cross-session similarity retrieval
  • Auto context compression, 512-token budget
FIG. V

Harness runtime

Event-sourced execution that is traceable and replayable. SQLite session persistence, dual token/time budgets, checkpoint resume, and hot-swappable modules.

  • Tool confirmation on dangerous ops
  • Checkpoint & resume on interrupt
  • Hot-plug with rollback
§ 03
  1. 01

    Web search

    Multi-query fusion across sources, body extraction, scoring, and source evaluation for deep research.

    SEARCH
  2. 02

    File operations

    Read / write / grep / glob with cross-directory browsing — “the XXX on my desktop” just works.

    FS
  3. 03

    Code execution

    Python & Bash with full terminal access — pipes, redirect, && / ||, env — under timeout isolation.

    EXEC
  4. 04

    Integrations

    GitHub search, PDF reading, weather query, web fetching, and a standalone Telegram bot.

    NET
§ 04 Out of the box
CommandAction
/deep_researchMulti-angle search → HTML report
!cmdExecute shell directly
@fileInject file contents into input
/modelSwitch model at runtime
/strategySwitch reasoning strategy
/compressManually compress context
/learnLearn a memory skill from chat
/memskillManage skills: list / evolve / snapshot
/sessionsBrowse historical sessions
/dashboardOpen the agent state panel
/draftsView stashed input drafts
Keys Ctrl+Sstash draft Ctrl+\dashboard ESCinterrupt long task ↑↓input history
§ 05 macOS · v1.0.0
RECOMMENDED Apple Silicon

LV Agent · arm64

.dmg · 101 MB · macOS 11+

Download .dmg ↓
INTEL x86_64

LV Agent · x64

.dmg · 105 MB · macOS 11+

Download .dmg ↓
ZIP Apple Silicon

LV Agent · arm64

.zip · 96 MB · portable

Download .zip ↓
ZIP x86_64

LV Agent · x64

.zip · 97 MB · portable

Download .zip ↓

Prefer the terminal? git clone https://github.com/Xinchen1/LV-Agent.git then ./install.sh for the global lv command. Backends: DeepSeek · OpenAI · Anthropic · OpenRouter · Ollama (offline).

The hand that draws and the mind that reasons are the same instrument — sharpened from different ends. — studio note, plate VII