README

Universal Agent Skills

Production AI agent skills and tool protocols across major runtimes

Live Showcase (Portfolio)  •  Dedicated Portal  •  Attribution Matrix  •  Contributing Guide  •  Security  •  Privacy Policy

Overview

Modern language models frequently generate generic conversational filler, speculative code abstractions, and hallucinated library methods. This repository delivers specialized, battle tested skill protocols that enforce senior engineering rigor, verified web intelligence, anti generic design aesthetics, and epistemic calibration.

Every skill follows the Universal Agent Skill specification, ensuring seamless interoperability across:

• Anthropic Claude Code and Claude Desktop
• Cursor IDE
• Google Antigravity and Gemini CLI
• OpenAI ChatGPT Custom GPTs and Codex
• OpenClaw and Windsurf
• Ollama, LangChain, and Local Models

Target Audience: Useful for anyone who wants to empower their AI agents or tools with production capabilities without technical friction, from independent creators to senior software engineers.

Quick Start and Installation

Install any skill into your preferred AI agent environment using the following standard workflows:

Model Context Protocol Server and Tools

Universal Agent Skills exposes twelve specialized tools through the Model Context Protocol for codebase knowledge graphs, document conversion, and multi network research.

Installation and Server Execution

To install and run the Model Context Protocol servers locally:

# Clone the repository
git clone https://github.com/mehanshbarthwal-lab/universal-agent-skills.git
cd universal-agent-skills

# Install dependencies for Python tool servers
pip install -e skills/graphify
pip install -e skills/markitdown/packages/markitdown
pip install -e skills/markitdown/packages/markitdown-mcp
pip install -e skills/agent-reach

# Run Graphify server over standard input and output
python -m graphify.serve path/to/graphify-out/graph.json

# Or run Graphify server over Streamable HTTP transport
python -m graphify.serve path/to/graphify-out/graph.json --transport http --port 8000

# Run MarkItDown server
python -m markitdown_mcp

# Run Agent Reach server
python -m agent_reach.integrations.mcp_server

Twelve Tools Catalog

1. get_status: Returns installation status, configuration health, and connectivity readiness across all research channels.
2. convert_to_markdown: Converts documents from web, local filesystem, or data URIs into structured Markdown.
3. query_graph: Queries the knowledge graph using BFS or DFS traversal and returns structural context.
4. get_node: Retrieves full attribute details and relational connections for a specified node label or identifier.
5. get_neighbors: Returns all direct neighbors and relational edges for a given node.
6. get_community: Retrieves all nodes belonging to a designated community cluster.
7. god_nodes: Computes degree centrality ranking to surface the most connected core abstractions.
8. graph_stats: Generates summary metrics covering node count, edge count, and community distributions.
9. shortest_path: Computes the shortest traversal path connecting two concepts in the knowledge graph.
10. list_prs: Lists open pull requests with CI status, review state, and impacted graph communities.
11. get_pr_impact: Analyzes blast radius for a given pull request by mapping changed files to graph communities.
12. triage_prs: Ranks actionable pull requests by review priority and community conflict merge risk.

Tool Annotations Specification

Every tool provides explicit boolean annotations matching its operational behavior:

Network Access and External Hosts

• Local Graphify Tools: query_graph, get_node, get_neighbors, get_community, god_nodes, graph_stats, and shortest_path operate entirely in memory and make zero network calls.
• Pull Request Tools: list_prs, get_pr_impact, and triage_prs make read only queries to api.github.com or invoke the local gh CLI.
• Document Conversion: convert_to_markdown queries remote HTTP and HTTPS hosts only when the user requests an external web URL. Loopback addresses, private IP ranges, and link local addresses are blocked by default.
• Research Channels: agent-reach communicates only with the specific host requested during active user searches (such as Reddit, GitHub, Bilibili, YouTube, or V2EX).

Local Files Read and Written

• Files Read: Local graph data files such as graphify-out/graph.json or custom project files; local document paths passed to convert_to_markdown (sensitive operating system paths such as private keys, shadow files, and credential stores are blocked); local configuration files in standard configuration directories.
• Files Written: All twelve tools are marked read only and do not write to or mutate project files. Graphify export routines write files only when explicitly requested through CLI options. Temporary scripts and scratch files are cleaned up immediately.

Credentials and Sensitive Files

Universal Agent Skills adheres to credential isolation and safety principles:

• Zero Read of Legacy Credentials File: The repository codebase never reads ~/.config/bird/credentials.env. That legacy file is written only when a user explicitly runs agent-reach config twitter-cookies --sync-legacy-twitter to enable backwards compatibility with third party scripts, created with restrictive 0o600 permissions. The uninstallation routine inspects whether the file exists solely to advise the user to perform manual deletion if desired.
• No Access to Sensitive Paths: The codebase never reads hardcoded absolute filesystem paths, external user profile paths, or sensitive system directories. All local configuration or input paths are resolved relative to the active workspace directory or supplied explicitly through environment variables and command line options.
• Lazy In Memory Evaluation: All API credentials and authentication tokens are loaded lazily at the exact time their specific feature or upstream query executes. No credentials are read at import or server startup.
• Strict Privacy and No Logging: Credentials and tokens are never printed to terminal output, never written to log files, never included in error messages, and never returned in Model Context Protocol tool responses. Credentials transmit exclusively to the intended service endpoint over encrypted TLS connections.

Environment Variables and Configuration

Variables used across the codebase are grouped into provider secrets, user settings, and ambient system detection. You only need to set a key if you actively use the corresponding feature.

Secrets and Provider Credentials (Group A)

User Configuration and Feature Flags (Group B)

Ambient System State (Group C)

Variables representing ambient system state (such as SSH_CONNECTION, SSH_CLIENT, DISPLAY, WAYLAND_DISPLAY, PYTEST_CURRENT_TEST, NVM_HOME, GH_CONFIG_DIR, XDG_CONFIG_HOME, NO_COLOR, CLAUDE_PROJECT_DIR, APPDATA, HOME, and TERM) are read strictly for environment detection, such as determining terminal color support or user configuration directory locations. These are not user configuration options and are never logged or stored.

Limitations

While Universal Agent Skills implements multi layer security checks, the following operational limitations apply:

• Network Level Egress: MarkItDown inspects initial destination URIs and every subsequent HTTP redirect against private, loopback, and link local IP blocks. However, protections implemented at the application layer cannot defend against advanced infrastructure threats such as DNS rebinding attacks where IP resolution shifts between validation and socket connection, or network tunnels bypassing application routing. For zero trust deployments, pair this server with operating system firewall egress filters or network proxies.
• Local Filesystem Sandboxing: File URI validation verifies resolved canonical realpaths and blocks recognized sensitive configuration markers and credential paths. Environments requiring strict multi tenant isolation should run Model Context Protocol servers within containerized namespaces or sandboxed virtual machines.

Authorship and Attribution

This repository places high value on open source transparency and honest provenance. Works are categorized into original engineering by Mehansh Barthwal, community upstreams, and methodology inspirations.

Original Engineering by Mehansh Barthwal

• Zero Hallucination Coder: Disciplined Discuss, Map, Decompose, Execute, Verify loop eliminating fabricated APIs. Officially merged into the community repository alirezarezvani/claude-skills (Pull Request #870).
• Universal Scraping Architect: Multi strategy web extraction framework balancing Firecrawl, Python, and hybrid pipelines with token budgeting and atomic checkpointing.
• Pain Point Miner: Consumer problem extraction system capturing real customer complaints from forums and review threads with mechanical deduplication.
• Loop Until Done: Autonomous self evaluation loop rewriting drafts until all objective rubric criteria clear.
• JD to Job: Career matching engine translating complex job descriptions into targeted operational qualifications.
• Resume Unrejectable: Four stage recruitment pipeline simulating ATS parsers, recruiter scanning timing, and executive depth.
• Locality Delivery Scraper: End to end scraping playbook for mapping locality food delivery and grocery markets.
• IPYNB Editor: Programmatic Jupyter Notebook cell inspection and patching utility preserving JSON schema integrity.
• Markdown Converter Router: Dynamic architectural router directing documents across AnyDoc and MarkItDown.
• Agent Meter Doctor: Universal agent telemetry, token waste auditing, TypeSafe Jev System One semantic friction diagnostics, dormant skill classification, and session context pruning paired with an autonomous post turn learning loop.
• Jev Decide: Sub second typed decision primitive powered by TypeSafe AI Jev model. Offloads binary conditions, categorical selections, and ordered scale ratings to System One decision models, eliminating generative token waste while preserving quantitative confidence scoring.

Adapted Open Source Projects

• ArXivist: Multi agent pipeline translating scientific research papers into executable codebases and Jupyter validation notebooks. Agent skill created by Mehansh Barthwal based on the original repository and concept by qosi org (qosi-org/arxivist).
• Humanizer: Authored by Siqi Chen (MIT License). Cleans machine writing hallmarks using Wikipedia style guides.
• Stop Slop: Authored by Hardik Pandya (MIT License). Blocks formulaic transitions and robotic filler during initial generation.
• Ponytail: Authored by DietrichGebert (MIT License). Senior developer minimalism suite auditing speculative abstractions.
• Graphify: Authored by Safi Shamsi (Apache 2.0 / MIT). Codebase knowledge graph generator producing relational maps and structural reports.
• GSD Core: Authored by Open GSD (MIT License). Spec driven autonomous development framework featuring phased execution gates.
• Agent Reach: Authored by Agent Eyes (MIT License). Multi platform research CLI and Model Context Protocol server for fifteen networks.
• Agent Video: Authored by Bradley Bonanno (MIT License). Media intelligence tool combining yt dlp, FFmpeg, and Whisper.
• Frontend Slides: Authored by Zara Zhang (MIT License). Interactive web presentation builder and PowerPoint to HTML slide converter.
• Ralph: Authored by snarktank (MIT License). Autonomous PRD conversion and implementation system.
• SkillOpt: Developed by Microsoft Corporation (MIT License). Agent self optimization and offline memory consolidation loop.
• MarkItDown: Utility by Microsoft Corporation (MIT License). Python library for multi format document extraction.
• Beautify GitHub Profile: Authored by Reza Shakeri (MIT License). Publication grade repository README documentation standard with badge templates and visual showcases.

Methodology Inspirations

• LLM Council: Multi persona deliberation methodology inspired by Andrej Karpathy.
• Karpathy Guidelines: Four lightweight coding instincts distilled from observations by Andrej Karpathy.
• Taste Skill: Modern anti slop frontend design guidelines enforcing typographic contrast.
• Truth Prompt: Epistemic calibration protocol separating verified fact from speculative inference.

Refer to ATTRIBUTIONS.md for full license notices, upstream URLs, and maintainer details.

Skill Catalog

1. Autonomous Coding and Software Engineering

2. Web Intelligence, Scraping, and Research

3. Writing, Editing, and Anti Slop Protocols

4. Orchestration, Planning, and Autonomous Pipelines

5. Visual Design and Presentation Engineering

6. Career Intelligence and Operational Positioning

7. Systems Architecture and Knowledge Management

Contributing

Contributions, new skill specifications, and bug reports are warmly welcomed. Please read CONTRIBUTING.md for full architectural guidelines, prompt engineering criteria, and attribution requirements before opening a pull request.

License

All original skill specifications and documentation authored by Mehansh Barthwal are released under the MIT License. Upstream tools and adapted suites retain their respective original licenses as documented in ATTRIBUTIONS.md.