The GitHub trending page this week tells a clear story: the tools ecosystem around AI coding agents is maturing fast. Google published a formal specification for describing visual design systems to AI agents. A new code intelligence engine indexes entire repositories into knowledge graphs and serves them over MCP. AWS released its official toolkit for letting AI agents interact with cloud infrastructure. Alibaba shipped an in-page JavaScript agent that controls web interfaces with natural language. And Cognee offers a self-hosted memory platform that gives AI agents persistent knowledge across sessions.
Here are five repositories worth your attention this week.
1. google-labs-code/design.md — A Design Spec File That AI Agents Can Actually Read
design.md from Google Labs is a format specification for describing a visual identity to coding agents. The idea is straightforward: you write a DESIGN.md file at the root of your project that combines machine-readable design tokens (YAML front matter) with human-readable design rationale (markdown prose). Agents that understand this format can produce UIs that faithfully follow your design system — exact colors, typography, spacing, component variants — without you pasting screenshots into every prompt.
The YAML front matter defines tokens like colors, typography, spacing, and component styles. The markdown body explains why those values exist and how to apply them. A token reference like {colors.primary} lets components reuse values without duplication. The spec supports any CSS color format including oklch(), and includes a version field currently set to alpha.
The project ships a CLI for linting and diffing DESIGN.md files. The linter checks for broken token references and validates WCAG contrast ratios, outputting structured JSON that agents can act on directly. The diff command detects token-level and prose regressions between two versions of a design system.
# Install and lint
npm install @google/design.md
npx @google/design.md lint DESIGN.md
# Compare two versions of a design system
npx @google/design.md diff DESIGN.md DESIGN-v2.mdIf you build frontends with AI coding agents and want consistent visual output without manual prompting, this spec gives agents the structured context they need. It’s still in alpha, but the format is clean and the tooling is already practical.
2. DeusData/codebase-memory-mcp — Index Your Entire Codebase as a Knowledge Graph
codebase-memory-mcp is a high-performance code intelligence MCP server that indexes codebases into a persistent knowledge graph. Written in pure C with zero dependencies, it ships as a single static binary for macOS, Linux, and Windows. The indexing speed is notable: an average repository takes milliseconds, and the Linux kernel (28 million lines of code, 75,000 files) completes in 3 minutes.
The tool uses tree-sitter AST analysis to parse 158 languages, enhanced with a Hybrid LSP semantic type resolution layer for Python, TypeScript, Go, Rust, and 7 more languages. It builds a graph of functions, classes, call chains, HTTP routes, and cross-service links — all stored in an in-memory SQLite database with LZ4 compression. Structural queries return results in under 1ms.
The research behind the project claims 83% answer quality with 10x fewer tokens and 2.1x fewer tool calls compared to file-by-file exploration, evaluated across 31 real-world repositories. It provides 14 MCP tools covering architecture overview, dead code detection, Cypher-like graph queries, and cross-service HTTP linking.
Installation is one command — the install script auto-detects Claude Code, Codex CLI, Gemini CLI, Zed, and 7 other agents, configuring MCP entries and instruction files for each. It also supports a team-shared graph artifact: commit a compressed snapshot to your repo, and teammates skip the reindex on clone.
# One-line install (macOS/Linux)
curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash
# Restart your coding agent, then say: "Index this project"For teams working with large codebases and AI coding agents, this replaces dozens of grep-and-read cycles with a single graph query. The zero-dependency, single-binary distribution makes it straightforward to adopt.
3. aws/agent-toolkit-for-aws — AWS’s Official Toolkit for AI Coding Agents
Agent Toolkit for AWS is the official, AWS-supported toolkit that gives AI coding agents the tools and guardrails to build, deploy, and manage applications on AWS. It reached general availability and bundles MCP servers, agent skills, and plugins that work with Claude Code, Codex CLI, Cursor, and Kiro out of the box.
The toolkit includes four plugin packages. aws-core covers service selection, CDK and CloudFormation, serverless, containers, storage, observability, billing, SDK usage, and deployment. aws-agents targets building AI agents on AWS with Amazon Bedrock and AgentCore. aws-data-analytics handles data lake, analytics, and ETL workflows with S3 Tables, AWS Glue, and Athena. aws-agents-for-devsecops integrates AWS DevOps Agent and AWS Security Agent for incident investigation, code review, vulnerability scanning, and penetration testing.
The core of the toolkit is the AWS MCP Server, a managed endpoint that gives agents access to the full AWS API — over 300 services through a single authenticated connection. It supports sandboxed Python script execution for complex multi-step operations and real-time documentation access. Enterprise controls include CloudWatch metrics, IAM condition keys that distinguish agent actions from human actions, and CloudTrail audit logging.
# Claude Code — install the core AWS plugin
/plugin install aws-core@claude-plugins-official
# Claude Code — install the AI agent-building plugin
/plugin install aws-agents@claude-plugins-officialIf you develop on AWS and use AI coding agents, this replaces ad-hoc prompt engineering about AWS services with structured, auditable tool access. The IAM condition keys that differentiate agent vs. human actions are particularly useful for governance.
4. alibaba/page-agent — Control Web Interfaces with Natural Language, In-Page
Page Agent from Alibaba is a JavaScript library that runs inside your web page and lets you control the interface with natural language. Unlike browser automation tools that require headless browsers or extensions, Page Agent works entirely in-page — no Python, no special permissions, no multi-modal LLMs.
The library processes the DOM as text, converts it into a structured representation, and sends it to an LLM that decides what actions to take — clicking buttons, filling forms, navigating menus. It works with any LLM through the standard OpenAI-compatible API format. The integration is minimal: a single script tag or an npm install, then instantiate a PageAgent with your model credentials.
import { PageAgent } from 'page-agent'
const agent = new PageAgent({
model: 'qwen3.5-plus',
baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
apiKey: 'YOUR_API_KEY',
language: 'en-US',
})
// Execute a natural language command against the current page
await agent.execute('Click the login button')
await agent.execute('Fill in the search box with "kubernetes monitoring"')The practical use cases are compelling. You can ship an AI copilot into any SaaS product without rewriting the backend — just drop in the script. Complex admin workflows with dozens of clicks collapse into a single sentence. An optional Chrome extension extends the agent across browser tabs for multi-page tasks, and a beta MCP server lets external agent clients control the browser.
If you build internal tools or complex web applications, Page Agent is worth evaluating as an accessibility and automation layer that requires no infrastructure beyond what’s already in the browser.
5. topoteretes/cognee — Persistent Long-Term Memory for AI Agents
Cognee is an open-source AI memory platform that gives agents persistent long-term memory across sessions. It ingests data in any format, builds a self-hosted knowledge graph, and lets agents recall and act on that knowledge with full context. The project combines vector embeddings, graph reasoning, and ontology generation to make documents both searchable by meaning and connected by evolving relationships.
The API is built around four operations: remember (store in the knowledge graph), recall (query with auto-routed search), forget (delete), and improve (refine knowledge). You can store information permanently in the graph or in session memory — a fast cache that syncs to the graph in the background. The recall operation auto-routes to the best search strategy, whether that’s vector similarity, graph traversal, or a hybrid approach.
import cognee
import asyncio
async def main():
# Store in the persistent knowledge graph
await cognee.remember("Our API uses JWT tokens with 15-minute expiry.")
# Store in session memory (fast cache, syncs to graph)
await cognee.remember("User prefers detailed error messages.", session_id="chat_1")
# Query with auto-routing to best search strategy
results = await cognee.recall("How does authentication work?")
for result in results:
print(result)
# Clean up when done
await cognee.forget(dataset="main_dataset")
asyncio.run(main())Cognee supports multiple LLM providers and ships prebuilt Docker images for both the API server and the MCP server. It offers a local UI, a CLI tool (cognee-cli), and integrations with Claude Code and OpenClaw. Client libraries are available in Python, TypeScript, and Rust. For teams building AI agents that need to maintain context across sessions — customer preferences, domain knowledge, organizational norms — Cognee provides the memory infrastructure without vendor lock-in.
Wrapping Up
This week’s trending repos reflect a shift from experimental AI agent tooling to production-grade infrastructure. design.md formalizes how agents understand visual systems. codebase-memory-mcp gives agents structured, efficient access to large codebases over MCP. Agent Toolkit for AWS brings official, auditable cloud access to the agent ecosystem. Page Agent proves that browser automation doesn’t require heavy infrastructure. And Cognee solves the persistent memory problem that every agent builder eventually hits.