AI Agent GitHub Digest — 2026-10-28

🌏 中文版

Today's Highlight

Two desktop apps on today's trending list are betting on the same future: a developer won't run just one agent, but five or ten at once, and the real problem becomes managing them from a phone. Orca and Cindy both chase that spot. Meanwhile wigolo wants to move web search off metered cloud APIs and back onto your own machine, and REA has already ridden "let the agent reverse-engineer someone else's feature for you" to 300,000+ stars. On the framework side, pydantic-ai shipped a release with a long compatibility-notes section worth reading before you upgrade.

Trending Repos

stablyai/orca ⭐ 107,899

GitHub · TypeScript · MIT

• What it is: A desktop app for running several coding agents at once — Claude Code, Codex, Grok, and more — each in its own isolated git worktree, with results compared side by side in one window. A mobile companion app lets you check progress, get notified when an agent finishes, and send a follow-up from your phone.
• Why it matters: "Fan one prompt across five agents, each in its own worktree, then merge the winner" used to mean hand-rolling scripts to manage worktrees and diff the results yourself. Orca turns that into a clickable UI, with GitHub and Linear task integration built in. Backed by YC, with a mobile layer on top, it's aimed squarely at the emerging problem of running too many agents at once to track by hand.
• Tech stack: Electron desktop shell + git worktree isolation + an embedded browser (Design Mode lets you click a UI element and drop it straight into a prompt) + iOS/Android companion apps
• Getting started: Easy — download the desktop build and sign in with your existing Claude Code or Codex subscription; prebuilt installers exist for macOS, Windows, and Linux

makecindy/cindy ⭐ 7,342

GitHub · TypeScript · Apache-2.0

• What it is: An open-source, cross-device AI agent client that wraps Claude Code, Codex, and its own lightweight agent "Pi" into one interface. You can pick up the same session on your phone that you started on your computer, without re-pasting the whole conversation.
• Why it matters: It isn't another agent framework — it's a shell that turns the coding agent you already use into something multi-device, with memory and scheduling. Your own computer becomes the server, so there's no separate box to rent; the mobile app also ships a built-in remote desktop for when a chat window isn't enough. For anyone already on Claude Code or Codex who wants a phone to step in, it's a ready-made upgrade path.
• Tech stack: Electron desktop client + Expo/React Native mobile client + a pnpm monorepo of shared packages (auth, device linking, and agent orchestration each live separately)
• Getting started: Easy — download the client and sign in with an existing subscription; to run a fully local agent offline, choose "Skip Sign-In," though that also drops the server-backed features (cloud sync, memory, skills)

KnockOutEZ/wigolo ⭐ 5,507

GitHub · TypeScript · AGPL-3.0

• What it is: A local-first MCP server for coding agents that bundles ten tools — search, fetch, crawl, structured extraction, cache, find-similar, and an autonomous research loop — all running on your own machine, with no API key required.
• Why it matters: It's worth comparing against metered search APIs like Firecrawl, Exa, and Tavily: wigolo uses a local browser engine plus on-device embedding/reranker models to make queries free and re-queryable offline, and its README includes a live four-way comparison against those three cloud services on the same question — failed engines and weak results included, not hidden. For budget-constrained setups, or agents that need to query frequently, it's a genuinely money-saving alternative.
• Tech stack: A Node.js MCP server + rank fusion across 18 search engines + local embedding/reranker models (needs roughly 1.5GB of disk)
• Getting started: Easy — npx wigolo init installs the browser engine and models in one command; --agents=claude-code,cursor wires up your everyday agents in the same run

morluto/rea ⭐ 396,382

GitHub · TypeScript · MIT

• What it is: Connects a coding agent to Hopper, Ghidra, IDA, and other reverse-engineering tools over MCP, so it can explain how a feature in an app works without seeing its source code — across dozens of formats from Electron, .NET, and Android APKs to WASM and GPU code objects.
• Why it matters: Most "understand this codebase" tools analyze your own repo. REA does the opposite: you see a feature you like in someone else's app, and the agent inspects it, gathers evidence, and builds you a version of its own. The case studies in its README — reconstructing DX-Ball's sound-pan calculation, tracing Notion's Electron clipboard bridge — make the "inspect → evidence → rebuild" pipeline concrete. Crossing nearly 400,000 stars within about six months of launch says this is a real, widely felt itch.
• Tech stack: A TypeScript CLI/MCP server + bridges into Hopper, Ghidra, and IDA + dedicated static parsers for dozens of file formats
• Getting started: Medium — npx rea-agents setup registers the agent in a few minutes, but deep native analysis still needs Hopper, Ghidra, or IDA installed first

Notable Releases

pydantic-ai v2.62.0

Release Notes

• Key changes: An agent can now change its thinking effort mid-conversation (via OpenAI's configuration_update or Anthropic's per-message output_config); three new model error categories — ModelRequestRejectedError, ModelAuthenticationError, and ModelNotFoundError — let an agent decide whether to retry or give up; hosted MCP tool names are now prefixed with the service by default (configurable via tool_prefix) to stop tool names from different MCP servers from colliding.
• Breaking changes: Yes — SummarizingCompaction's summary is now stored as a CompactionPart instead of being folded into the system prompt; GraphBuilder now starts broadcast branches in .to() order instead of reversed, which can change replay behavior for durable workflows that were paused mid-queue before the upgrade; monty_sandbox_url and cache_ttl are both deprecated.
• What it means for you: If your pydantic-ai agent uses SummarizingCompaction for long-conversation compaction, or you're upgrading across versions with durable workflows still queued, read this release's compatibility notes first — especially the replay-order change.

browser-use 0.13.14

Release Notes

• Key changes: Added OpenAI browser tools integration; fixed screenshot, crop, and click coordinates misaligning on scaled (non-100%) displays; added support for Browser Use Cloud SDK V4.
• Breaking changes: None.
• What it means for you: If you've been running browser-use automation on a laptop at its native (often scaled) resolution and clicks kept landing in the wrong place, this release should noticeably fix that.

Today's Takeaway

I assumed multi-agent collaboration would mature first at the framework layer — orchestration code inside something like LangGraph or CrewAI. Watching Orca and Cindy today flipped that: the layer that's actually maturing first is the consumer interface. Nobody wants to write orchestration code; they want a desktop app that opens five worktrees for them and a phone that pings when one is done. Frameworks are solving how agents talk to each other. These desktop tools are solving how a single person stays on top of a fleet of them. Two different problems, both equally urgent.

References

• stablyai/orca — GitHub repo
• makecindy/cindy — GitHub repo
• KnockOutEZ/wigolo — GitHub repo
• morluto/rea — GitHub repo
• pydantic-ai v2.62.0 Release Notes
• browser-use 0.13.14 Release Notes
• Top 10 AI GitHub Repositories - 2026-10-27 — cross-check source for today's trending list
• GitHub Weekly Trending Snapshot on October 27, 2026 | GitStock — cross-check source for today's trending list