AI Open Source Radar — October 28, 2026

Method: GitHub Trending bot issue happydog-intj/ai-xiaohongshu-daily #602 (Morning edition, Oct 28, Top 9) compared with the previous day’s #599 (Oct 27, 02:00 slot, Top 9). Primary metric is stars today. Excluded: repos above 40k total stars, long-running chart regulars, slowing repos, non-AI repos and perennial giants. All four repos featured today grew faster than yesterday, so all return.

Nutlope/hallmark — +2,859 ⭐ today (3 days in a row)

URL: https://github.com/Nutlope/hallmark
Total stars: 38,591 · Day over day: +2,235 → +2,859 (+28%) · Language: CSS · License: MIT

Hallmark is a design skill from Together AI for Claude Code, Cursor and Codex that pushes coding agents away from generic, obviously AI-made UI toward intentional design. At its core is a SKILL.md rule set with reference docs: it first picks a page structure that fits the brief, applies one of 21 visual themes, then runs 57 slop-test gates and a self-critique pass before output to strip out common AI design defaults. Commands:

Default (build): generate new UI under the rule set, then verify it
hallmark audit <target>: score existing code against anti-patterns and return a fix list
hallmark redesign <target>: rebuild structure while keeping copy, information architecture and brand
hallmark study <screenshot|URL>: extract structure, type pairing and color from a reference into design.md

Install with a single line: npx skills add nutlope/hallmark. Daily gains have grown three days straight (+1,359 → +2,235 → +2,859), and it again tops all non-excluded repos. At 38.6k stars it is just short of the 40k cutoff, so it will likely drop out next time.

Practical use: Run audit or redesign on AI-built landing pages and internal tools as a design QA step that removes the template look.

#AgentSkills #FrontendDesign #ClaudeCode #DesignSystems

makecindy/cindy — +1,446 ⭐ today (4 days in a row)

URL: https://github.com/makecindy/cindy
Total stars: 7,339 · Day over day: +1,268 → +1,446 (+14%) · Language: TypeScript · License: Apache-2.0

Cindy is an open-source local AI agent that works out of the box, operating on your real files and signed-in apps and controlling the browser, computer and phone. It is a pnpm monorepo with an Electron desktop app, an Expo/React Native mobile app, and shared packages for auth, device pairing, agent orchestration and model providers, and it bundles the Claude Code and Codex binaries as harnesses. Key features:

Switch models or harnesses mid-task while keeping workspace, memory, skills and tools
Split work in parallel and have agents with different harness and model combos cross-review results
Memory, skills, scheduled automations, MCP integrations and plugins
Choose the official cloud, an existing Claude Code or Codex subscription, your own API keys, or local models

Daily gains have risen four days in a row (+471 → +677 → +1,268 → +1,446), pointing to steady adoption rather than a one-off spike. Its pitch as a unified desktop agent that reuses existing coding-agent subscriptions keeps drawing interest.

Practical use: Teams already paying for Claude Code or Codex can trial desktop and mobile automation plus cross-review workflows at no extra cost.

#DesktopAgent #MultiAgentOrchestration #MCP #ComputerUse

JetBrains/thinkrail — +1,211 ⭐ today (2 days in a row)

URL: https://github.com/JetBrains/thinkrail
Total stars: 3,801 · Day over day: +1,156 → +1,211 (+5%) · Language: TypeScript · License: Apache-2.0

ThinkRail is an open-source agentic IDE from JetBrains built around the pi coding agent, running as a desktop app or in a browser launched by the thinkrail CLI. A Bun-based HTTP/WebSocket engine host runs one pi session per chat and talks to a React 19, Zustand and Tailwind v4 UI over a typed, versioned protocol called wire. pi handles models, skills, compaction and cost, while ThinkRail owns the workspace and editor. Key features:

A separate git worktree and branch per workspace, so multiple agents run in parallel
Push, open PRs, check CI and merge from inside the app
Monaco editor, file tree with git change markers, terminal, and review comments sent to chat
Spec-driven development: agents read and update a SPEC.md graph next to the code via spec_* tools
Reuses existing Agent Skills and MCP servers (mcp.json), with JetBrains AI or any provider configured in pi

After its first-day surge (+52 → +1,156), it held the pace on day two with +1,211, suggesting real user inflow beyond launch buzz. It reached 3.8k stars in two days.

Practical use: Fits team workflows that run agents on separate branches and compare or merge results, and experiments in spec-driven development where spec docs persist across sessions.

#AgenticIDE #GitWorktree #SpecDrivenDevelopment #JetBrains #VibeCoding

nightly-labs/openbot — +794 ⭐ today (2 days in a row)

URL: https://github.com/nightly-labs/openbot
Total stars: 2,215 · Day over day: +312 → +794 (+154%) · Language: TypeScript · License: PolyForm Noncommercial 1.0.0 (Apache-2.0 through v0.1.11)

OpenBot is a local-first desktop workspace that runs local CLIs such as Codex, Claude and Grok (plus OpenCode, Gemini, Cursor CLI and Cline via ACP) as always-on AI teammates. The Electron main process launches provider processes over stdio and streaming JSON and records them in a SQLite event log, with a SolidJS UI connected through typed IPC. Team connections use WebRTC, and the account server (Cloudflare D1 and R2) handles accounts only. Key features:

Independent session and workspace per agent with automatic context compaction
FIFO message queue with pause, resume and cancel that survives crashes
Agent-to-agent messages and file handoffs, plus explicit task delegation in shared channels
Built-in browser that agents control, with optional Computer Use
Remote hosting on Docker or Linux servers, an iOS app and webhook triggers

Its daily gain grew 2.5x overnight (+312 → +794), making it No. 2 on today’s list behind hallmark. It is still a developer preview in which agents use file, command and network permissions without confirmation, and the latest version is under a noncommercial license.

Practical use: Good for experimenting with running coding agents from several providers side by side as teammates, but commercial use requires a separate license.

#AITeammates #LocalFirst #AgentToAgent #ComputerUse

Excluded: obra/superpowers (+608, 306.2k), pbakaus/impeccable (+1,458, 90.3k) and affaan-m/ECC (+1,183, 290.6k) are above 40k stars. gitroomhq/postiz-app (+953, 39.2k) is a social scheduling tool that has trended repeatedly since 2024 (Trendshift badges not verified). togg53192-cmd/jailbreaks (+890) is a list of model jailbreak prompts, not a practical tool. Yesterday’s TradingAgents, SkillSpector and Agent-Reach are absent from today’s list.