Skip to content

README.md

Latest commit

 

History

History
162 lines (112 loc) · 7.81 KB

README.md

File metadata and controls

162 lines (112 loc) · 7.81 KB
Raw

LabKits

LabKits logo

Research operating kits for AI agents, built by Mark at Mindify AI.

LabKits is the researcher-facing sibling of ideas like gStack or gBrain: a portable stack of skills, prompts, harnesses, and environment conventions that help an AI agent work like a serious research partner instead of a generic chatbot. The first kits include Cothinking for active paper reading, Math Knowledge Graph for dense formal reasoning, Agentic Solution Search for searching solution spaces, LaTeX Paper Writing for venue-ready manuscripts, Fine-Tune Hyperparameter Search for reproducible ML tuning, and Hypothesis Experiment Loop for designing experiments that either meet a hypothesis or force a principled revision.

Install

macOS, Linux, and WSL:

curl -fsSL https://raw.githubusercontent.com/mindify-ai/LabKits/main/scripts/install.sh | bash

Windows PowerShell:

irm https://raw.githubusercontent.com/mindify-ai/LabKits/main/scripts/install.ps1 | iex

From a cloned checkout:

./scripts/install.sh

The installer creates ~/.labkits, installs the packaged skill archives, installs the generic agent harness, and natively installs each skill into every detected harness: ~/.codex/skills for Codex, ~/.claude/skills for Claude Code, and a Gemini CLI extension (with a generated gemini-extension.json and GEMINI.md) under ~/.gemini/extensions. Set LABKITS_INSTALL_CODEX, LABKITS_INSTALL_CLAUDE, or LABKITS_INSTALL_GEMINI to "0" to skip any of them; see env/labkits.env.example for the full list of overridable env vars.

What Is Included

  • skills/paper-reading/cothinking.skill - packaged Cothinking skill archive.
  • skills/math-understanding/math-knowledge-graph.skill - packaged Math Knowledge Graph skill archive.
  • skills/agentic-simulation/agentic-solution-search.skill - packaged agentic solution-search skill archive.
  • skills/experiment-design/hypothesis-experiment-loop.skill - packaged hypothesis-driven experiment-loop skill archive.
  • skills/paper-writing/latex-paper-writing.skill - packaged LaTeX paper-writing and venue-readiness skill archive.
  • skills/ml-training/fine-tune-hyperparameter-search.skill - packaged ML fine-tuning and hyperparameter-search skill archive.
  • skills/*/*/ - editable source for each LabKits skill.
  • agent-harness/LABKITS.md - environment-independent harness instructions for agents that do not natively load .skill files.
  • scripts/install.sh and scripts/install.ps1 - one-command installers for Unix-like shells and Windows PowerShell.
  • scripts/doctor.sh - local environment and package validation.
  • scripts/package-skill.sh - rebuilds the .skill archive from source.
  • env/labkits.env.example - optional environment variables for predictable installs.
  • CONTRIBUTE.md, SECURITY.md, and CHANGELOG.md - OSS project guidance.

Why LabKits

Most AI research help stops at summarization. LabKits pushes the agent into a more useful role:

  • Active reading: the agent asks the researcher to predict, critique, and transfer ideas before revealing answers.
  • Math decomposition: the agent turns long formal statements into assumptions, definitions, equations, transformations, claims, and reasoning-flow edges.
  • Agentic simulation: the agent maps a solution space, keeps a frontier, runs cheap probes, prunes failed paths, and verifies convergence.
  • Experiment design: the agent derives predictions, designs discriminating tests, optimizes within a hypothesis, and revises the hypothesis from evidence.
  • Paper writing: the agent drafts, revises, audits, and preflights LaTeX manuscripts against current venue instructions.
  • Model tuning: the agent designs disciplined fine-tuning sweeps, tracks trials, prunes weak configs, and packages reproducible recipes.
  • Visual compression: concepts become Mermaid graphs, ASCII sketches, figure walkthroughs, or small study companions when useful.
  • Research taste: the agent separates claims, hypotheses, evidence, and weak links.
  • Portable setup: the same kit can be used inside Codex, other agent runtimes, or a plain prompt harness.

Think of it as a small research lab bench you can carry between machines and agent environments.

Use

In Codex or any agent that can load installed skills, ask for Cothinking explicitly:

Use LabKits Cothinking to walk me through this paper: https://arxiv.org/abs/...

Or unpack dense math directly:

Use LabKits Math Knowledge Graph to explain this loss function at undergraduate and expert levels.

Or search an open solution space:

Use LabKits Agentic Solution Search to explore possible approaches before choosing an implementation.

Or prepare a manuscript:

Use LabKits LaTeX Paper Writing to convert these results into an ICML-ready paper draft.

Or tune a model:

Use LabKits Fine-Tune Hyperparameter Search to design a LoRA sweep for this dataset.

Or test a hypothesis:

Use LabKits Hypothesis Experiment Loop to design experiments for this hypothesis and revise it if results contradict it.

For agents without native skill loading, paste or attach agent-harness/LABKITS.md as system or project context, then provide the paper.

OS And Agent Environment

Environment Status Notes
macOS Supported Uses bash, curl, and unzip; installs to ~/.labkits and ~/.codex/skills when available.
Linux Supported Same path and tool requirements as macOS.
WSL Supported Treated as Linux; recommended for Windows users who run Unix-style agents.
Windows PowerShell Supported Uses Invoke-RestMethod and Expand-Archive; installs to $HOME\\.labkits.
Codex Supported Skill source is expanded into ~/.codex/skills.
Claude Code Supported Skill source is expanded into ~/.claude/skills, using the same SKILL.md the other harnesses read.
Gemini CLI Supported Each skill is installed as a Gemini CLI extension under ~/.gemini/extensions, with a generated gemini-extension.json manifest and GEMINI.md context file.
Generic agents Supported Use agent-harness/LABKITS.md as the portable harness.

Every skill also ships an agents/<harness>.yaml interface file (openai.yaml, claude.yaml, gemini.yaml) describing its display name, short description, and a default invocation prompt for that harness. See "Add A Harness" in CONTRIBUTE.md for how to extend install support to another harness.

Optional environment variables (see env/labkits.env.example for the full list):

export LABKITS_HOME="$HOME/.labkits"
export LABKITS_REPO_RAW="https://raw.githubusercontent.com/mindify-ai/LabKits/main"

export CODEX_HOME="$HOME/.codex"
export LABKITS_INSTALL_CODEX="1"   # set to "0" to skip Codex

export CLAUDE_HOME="$HOME/.claude"
export LABKITS_INSTALL_CLAUDE="1"  # set to "0" to skip Claude Code

export GEMINI_HOME="$HOME/.gemini"
export LABKITS_INSTALL_GEMINI="1"  # set to "0" to skip Gemini CLI

Verify

./scripts/doctor.sh

The doctor checks OS support, required commands, archive integrity, source layout, per-harness interface files, and likely install state across all supported harnesses.

Develop

Edit the source skill under skills/<category>/<skill-name>/, then rebuild the packages:

./scripts/package-skill.sh
./scripts/doctor.sh

The package command rewrites each skills/<category>/<skill-name>.skill archive from its source directory.

Roadmap

  • More paper-reading kits for review writing, experiment reproduction, and related-work mapping.
  • A benchmark harness for testing whether a kit improves researcher comprehension or decision quality.
  • Runtime adapters for more agent environments (Cursor, Windsurf, and similar).
  • Versioned releases of individual kits.

License

MIT License. Copyright (c) 2026 Mark Chen.