Scaling GitHub for your Agents
summary
GitHub reports over 4 million downloads of the stdio server and says the remote MCP server serves about 7 million tool calls a week. Architecture is stateless and horizontally scalable, using Redis for session storage and building a fresh server instance per request. Tool proliferation is solved with tool sets and dynamic selection, yielding ~49% context reduction and >75% token output cuts in places, plus ~95% tool success after intent-driven tooling and OAuth 2.1 PKCE.
AI-written summary · about
chapters (13)
- 0:00Start
- 0:29Overview of GitHub's MCP public launch and community growth.
- 2:06Challenges of tool proliferation and impact on agent context.
- 3:21Mitigation via "tool sets" and dynamic discovery.
- 5:54Optimizing API output tokens to improve efficiency.
- 6:44Improving reliability through intent-based tool design.
- 8:14Security strategy: OAuth 2.1 and PKCE implementation.
- 10:40Managing prompt injection and security vulnerabilities.
- 12:35Using OAuth scopes for granular tool filtering.
- 13:47Stateless server architecture and Redis session management.
- 15:18Experimental features and human-in-the-loop UX.
- 16:30Future outlook: Compositional tools and automation.
- 18:04Final project metrics: Downloads, forks, and volume.
related talks
36mPutting AI Agents to Work: A New Era of Open Source Connectivity with MCP
45mHow GitHub builds GitHub on GitHub
31mGitHub as a self-sustaining product for large-scale environments
45mHow GitHub builds GitHub with GitHub
01h 06mHow GitHub delivers GitHub using GitHub
45mHow Github uses GitHub to develop GitHub
01h 03mAI Coding Agents and how to code them
18mShipping at Scale: ChatOps at GitHub #bbuzz