AI Agent GitHub Digest — 2026-08-30

🌏 中文版

Today's Highlights

Today's trending projects happen to all be patching up coding agents' ability to "perceive the outside world" — chrome-devtools-mcp lets an agent actually see what's happening inside the browser, GitNexus lets an agent understand how code relates to other code instead of guessing line by line, and context-mode tries to stop agents from burning a limited context window on tool output they can't parse. Framework-side changes were light this time; pydantic-ai's @durable_operation is the one architectural addition worth noting.

Trending Repos

ChromeDevTools/chrome-devtools-mcp ⭐ 50,162

GitHub · TypeScript · Apache-2.0

• What it is: an MCP server maintained by the Chrome DevTools team itself, letting coding agents like Claude, Cursor, and Copilot control and inspect a real Chrome instance through the MCP protocol — not simulated browser behavior, but a direct hookup to the DevTools Protocol itself.
• Why it's worth a look: most "AI controls a browser" tools stop at clicking, screenshotting, and reading the DOM. This server opens up DevTools' full performance profiling (trace recording plus actionable performance insights), network-request inspection, and console messages with source-mapped stack traces to the agent — effectively handing the agent the exact toolset a human engineer reaches for when debugging. Under the hood it uses puppeteer for automation and auto-waits for actions to settle, cutting down on misjudgments where "the agent clicked the button but the screen hasn't updated yet." It hit #1 on GitHub Trending's TypeScript chart today.
• Tech Stack: TypeScript + puppeteer + Chrome DevTools Protocol; officially only Google Chrome / Chrome for Testing are guaranteed to work
• Getting Started: Low — npx chrome-devtools-mcp@latest plugs into any MCP client. Note that it exposes the entire browser tab content to the MCP client, so it's not a good fit for driving tabs with sensitive data.

abhigyanpatwari/GitNexus ⭐ 46,399

GitHub · TypeScript · PolyForm Noncommercial

• What it is: a code knowledge-graph generator that runs entirely in the browser (no server needed). Point it at a git repo (GitHub, GitLab, Azure, or a local ZIP all work) and it builds an interactive graph covering dependencies, call chains, and module clusters, queryable through a built-in Graph RAG agent.
• Why it's worth a look: most "AI understands your codebase" tools work by chunking files into a vector database for semantic retrieval, which tends to miss structural relationships like "who calls this function, and what does it call in turn." GitNexus goes the other way — it builds the structural graph first, then lets the agent traverse it. The project positions itself as "one layer deeper than DeepWiki": DeepWiki helps you understand what the code says, GitNexus lets you analyze how the code is wired together. Its CLI + MCP modes plug directly into Cursor, Claude Code, and Codex, giving an agent an architectural view before it touches a large repo, reducing the risk of a change in one spot rippling into others it shouldn't.
• Tech Stack: TypeScript, runs entirely on the frontend (no backend dependency), available as either an MCP server or a web UI
• Getting Started: Medium — the CLI/MCP mode is quick to pick up, but getting the full benefit of graph analysis (especially on large repos) is easier after reading the docs on how the indexing pipeline works.

mksglu/context-mode ⭐ 20,243

GitHub · TypeScript · ELv2

• What it is: a middleware layer that reduces how much context window a coding agent wastes — it "sandboxes" a tool call's raw output before returning it to the model and preserves task memory across sessions; the official docs claim a large reduction in the token footprint of tool output.
• Why it's worth a look: today's coding agents commonly get their context blown out by something as simple as a single grep returning 3,000 lines, leaving the model with no budget left for the thinking that actually matters. context-mode chooses to compress at the tool-output layer instead of asking users to learn to write more precise queries, and via a hooks mechanism it claims to work uniformly across 17 agent platforms (Claude Code, Cursor, Codex, Copilot, and others). One caveat: the "adopted by Microsoft/Google/Meta and other enterprises" badge links in the README currently point nowhere — that's a self-reported claim with no independently verifiable source found so far, so weigh the adoption claim accordingly.
• Tech Stack: TypeScript, dual integration via MCP and hooks
• Getting Started: Low — per the official docs it installs via a package manager and hooks into an existing agent config without standing up your own server.

google/skills ⭐ 18,974

GitHub · Python · Apache-2.0

• What it is: Google's officially maintained Agent Skills package library, covering Google Cloud-related operational skills — everything from "how to authenticate into GCP" to "deploying an agent on GKE" and "building enterprise-grade RAG with AlloyDB." Run npx skills add google/skills to pick and install into any coding agent that supports the Agent Skills standard.
• Why it's worth a look: this fits the "skill ecosystem expansion" trend we've tracked since the 08/16 series, except this time it's a cloud vendor showing up directly — Google is repackaging its own product docs into agent-executable skill packages instead of leaving that to the community to reverse-engineer. For teams already wiring Google Cloud services into an agent workflow, this is a more reliable starting point than learning the docs from scratch on your own.
• Tech Stack: Markdown/YAML skill definitions plus the skills.sh install ecosystem
• Getting Started: Low — npx skills add google/skills to pick and install; still under active development (the README explicitly says it's not yet stable).

livekit/agents ⭐ 13,560

GitHub · Python · Apache-2.0

• What it is: LiveKit's real-time voice/video AI agent framework, handling streaming audio, multi-turn conversation state, and integrating different voice providers, so developers don't have to write their own real-time communication pipeline layer.
• Why it's worth a look: not a new project, but it's still landing commits today (updated earlier today), a sign this space is still iterating fast. Unlike a plain text chatbot, voice scenarios have to deal with far more complex interruption, latency, and streaming-transcription-sync problems, and this framework collapses all that low-level mess into a unified API — currently one of the most widely adopted open-source options in this category.
• Tech Stack: Python, with an abstraction layer that plugs into multiple voice/LLM providers including OpenAI and Deepgram
• Getting Started: Medium — the core concepts (room, track, pipeline) take a bit of time to internalize, but there are several official example projects to copy from as a starting point.

Notable Releases

pydantic-ai v2.36.0

Release Notes

• Key changes: adds the @durable_operation decorator, opening a public backend API that lets third-party durable-execution engines (beyond Prefect) plug in and take over long-running agent tasks; the clai CLI gains --mcp-config support and tool-call streaming.
• Breaking changes: @durable_operation now requires an explicit operation name — usage that previously omitted it and let the framework infer it automatically will now fail.
• What it means for you: if you're already using @durable_operation without an explicit operation name, add one before upgrading. If you're evaluating fault-tolerance for long-running agent tasks, this newly opened backend API means you're no longer locked into a single durable-execution vendor.

Today's Takeaway

I used to think "giving an agent better perception" mostly meant structured sources like the filesystem and databases. Seeing chrome-devtools-mcp open up the entire browser debugging toolchain made it click that the browser itself is becoming a first-class environment for agents — not just something automated tests execute against, but a debugging and information source on equal footing with the filesystem.

References

• ChromeDevTools/chrome-devtools-mcp
• chrome-devtools-mcp README
• abhigyanpatwari/GitNexus
• GitNexus README
• mksglu/context-mode
• context-mode README
• google/skills
• google/skills README
• livekit/agents
• pydantic-ai v2.36.0 Release Notes
• GitHub Trending — TypeScript (daily)
• GitHub Trending — Python (daily)