• AI Engineering
  • Posts
  • [Hands on] Build a GitHub Intelligence Agent with GitHub MCP

[Hands on] Build a GitHub Intelligence Agent with GitHub MCP

... PLUS: SkillNet: An Open-Source Skill Marketplace for AI Agents

Author

Sumanth P
May 22, 2026

In today’s newsletter:

  • [Hands on] Build a GitHub Intelligence Agent with GitHub MCP

  • SkillNet: An Open-Source Skill Marketplace for AI Agents

Reading time: 5 minutes.

You're evaluating an open-source library. You check the star count, skim the commit history, look at open issues, and hunt down the maintainer's profile to see what else they've shipped.

By the time you have a real picture, you have twelve browser tabs open, a half-finished notes doc, and you've spent 40 minutes on research that should have taken five.

GitHub's official MCP server can retrieve these informations. It wraps the full GitHub API as a callable tool catalog for any MCP-compatible agent.

Point an LLM at it, and GitHub research becomes a conversation. We built an agent that does exactly that.

Prerequisites

Tech Stack

  • Gemini 3 Flash (gemini-3-flash-preview) handles reasoning and tool calls

  • GitHub's official MCP server at api.githubcopilot.com/mcp/ provides the tool catalog

  • Haystack (Agent Framework)

  • Streamlit powers the UI

Connecting to GitHub

The MCP server lives at a single URL. You point the agent at it with a Personal Access Token and that's the entire connection setup:

What the agent knows to do

The system prompt defines the agent's scope and behavior:

Two things worth noting:

  1. The prompt explicitly tells the agent to retry with a different strategy if a tool call fails

  2. It ends with a hard instruction to skip narration and output only the final answer. Clean responses, no step-by-step thinking in the UI.

Putting it together

The streaming callback wires the agent's output directly to Streamlit so responses appear token by token as the agent works.

What you can research

Any question GitHub's API can answer:

  • "What has Andrej Karpathy been building lately?"

  • "Show me open issues in huggingface/transformers"

  • "Find trending Python repos for RAG pipelines"

  • "Who are the top contributors to pytorch/pytorch?"

  • "What did Microsoft ship on GitHub last week?"

The agent picks the right tools per query, fires the calls, and writes up the results. The Streamlit UI shows which tools were used and the elapsed time per response.

Running the agent

The app runs in Streamlit. Credentials go in the sidebar. No .env setup required. You need a Gemini API key from Google AI Studio and a GitHub PAT with read-only repo access. 

git clone https://github.com/Sumanth077/Hands-On-AI-Engineering.git
cd Hands-On-AI-Engineering/ai_agents/github_intelligence_agent
uv run streamlit run app.py

Open http://127.0.0.1:8501, paste your keys, and start querying.

Every time you deploy an agent to handle a task, it figures out how to do it from scratch.

There is no shared memory between agents, no way to hand off what one learned to another, and no community layer where solved problems get reused.

SkillNet is built to fix that.

What SkillNet Is

SkillNet is an open-source platform that treats AI agent skills as first-class, shareable packages, like npm for AI capabilities.

Instead of every agent reinventing the same workflows, skills get packaged, published, searched, and installed the same way software dependencies do.

The platform provides end-to-end tooling to search, install, create, evaluate, and organize skills, so agents can learn from the community and continuously grow. The knowledge base currently holds over 300,000 skills.

The Four Things You Can Do With It

Search. Find skills by keyword or by semantic query. Looking for something that handles financial PDF analysis isn't a keyword match, it's an intent match, and the vector search handles that.

Install. Pull any skill into your agent environment via the Python SDK or CLI. No API key needed for search and download.

Create. This is the interesting one. SkillNet can generate a structured skill package from almost any source:

  • A GitHub repository

  • A PDF, PowerPoint, or Word document

  • A conversation log or agent execution trace

  • A plain text prompt

That last two matter most. If your agent already solved a problem in a previous session, that trajectory can be distilled into a reusable skill. Experience converts directly into capability.

Evaluate. Skills go through multi-dimensional evaluation before they're trusted. Not every skill in the marketplace is equal, and the evaluation pipeline surfaces quality differences.

Why the Creation Pipeline Matters

Most skill systems require you to write skills by hand. SkillNet lets you extract them from what agents have already done.

An agent finishes a task. The execution trace, the sequence of steps it took to get there, gets passed to SkillNet.

SkillNet converts that into a structured skill package. The next agent that faces the same problem doesn't start from scratch. It installs the skill and picks up where the first one left off.

This is how the knowledge base grows without requiring manual authorship at scale. Agents generate experience, experience becomes skills, skills get shared.

Where It Plugs In

SkillNet has an MCP server, which means any MCP-compatible agent can access the skill marketplace without custom integration.

It's also natively built into OpenClaw, where agents search, download, and apply skills automatically as part of task execution. The platform is open-source under MIT license.

That’s all for today. Thank you for reading today’s edition. See you in the next issue with more AI Engineering insights.

PS: We curate this AI Engineering content for free, and your support means everything. If you find value in what you read, consider sharing it with a friend or two.

Your feedback is valuable: If there’s a topic you’re stuck on or curious about, reply to this email. We’re building this for you, and your feedback helps shape what we send.

WORK WITH US

Looking to promote your company, product, or service to 200K+ AI developers? Get in touch today by replying to this email.