This week's GitHub Trending list is not just another pile of starred repos. The interesting pattern is that builders are trying to give agents better memory, broader web reach, safer skills, richer sandboxes, and more serious creative workflows.
My favorite repo this week is DeusData/codebase-memory-mcp. It attacks one of the most expensive problems in AI coding: every new agent session has to rediscover the codebase. If a persistent code memory layer works well, it saves tokens, time, and mistakes.
Source Note
This roundup is based on GitHub Trending weekly, checked on June 22, 2026, plus the repo links and weekly star counts supplied in the brief. GitHub Trending changes quickly, so treat the star counts as a snapshot, not a permanent ranking.
JQ AI SYSTEMS is adding a practical builder/operator read: what each repo is useful for, where it fits in an AI workflow, and what to watch before installing it.
The Short Version
- Best overall: codebase-memory-mcp, because codebase memory is becoming core infrastructure for coding agents.
- Most powerful but risky: Agent-Reach, because internet reach is useful but permissions and platform rules matter.
- Best safety layer: SkillSpector, because agent skills are now supply-chain objects.
- Best creative workflow: OpenMontage, because video production is becoming agentic.
- Best process layer: agent-skills, because better engineering procedure beats faster chaos.
Weekly Trending Table
I would not rank these only by stars. Some high-star repos are broad public utilities. Some lower-star repos are much more relevant to AI agent workflows. The table below includes a GitHub link for every repo, plus the JQ AI SYSTEMS usefulness score.
| Repo | Weekly signal | What it does | JQ AI SYSTEMS read | Score |
|---|---|---|---|---|
| DeusData/codebase-memory-mcp | 6.3k+ this week | Persistent codebase memory for agents |
Indexes a codebase into a knowledge graph so coding agents can ask precise questions instead of repeatedly searching the same files. Watch: Great fit for large repos, but treat the index as infrastructure. Check what it stores and how your team handles private code. |
9.6/10 |
| Panniantong/Agent-Reach | 8.2k+ this week | Web reach for AI agents |
Aims to let agents read and search sources like X, Reddit, YouTube, GitHub, Bilibili, and XiaoHongShu without separate platform APIs. Watch: Use carefully. Web automation can collide with login rules, cookies, copyright, scraping limits, and platform terms. |
9.3/10 |
| NVIDIA/SkillSpector | 4.0k+ this week | Security scanner for AI agent skills |
Scans installable agent skills for risky behavior, malicious patterns, and security issues. Watch: A scanner is not a guarantee. Still review skills manually before giving them tools, files, network, or account access. |
9.2/10 |
| addyosmani/agent-skills | 5.6k+ this week | Production engineering skills for coding agents |
Pushes agents toward better software delivery habits: spec, plan, build, verify, review, simplify, and ship. Watch: Skills help only if the user lets them slow the work down. They are not magic if you still ask for one-shot chaos. |
9/10 |
| calesthio/OpenMontage | 2.8k+ this week | Agentic video production studio |
Packages video production into pipelines, tools, and agent skills so an AI coding assistant can help create video systems. Watch: Video automation still needs rights checks, editorial review, brand control, and cost controls for generation APIs. |
8.9/10 |
| withastro/flue | 1.2k+ this week | Sandbox agent framework |
Points toward a cleaner way to run agents inside controlled sandboxes instead of trusting them inside the main workspace. Watch: Early frameworks need testing against your own threat model, permissions, filesystem boundaries, and review process. |
8.7/10 |
| google-research/timesfm | 4.1k+ this week | Time-series foundation model |
Useful for teams forecasting demand, sales, operations, infrastructure load, inventory, or other time-series signals. Watch: Forecasts are not decisions. Pair with domain review, backtesting, and clear error bands. |
8.4/10 |
| LMCache/LMCache | Trending this week | KV cache layer for LLM systems |
Infrastructure for making repeated LLM workloads faster and cheaper by reusing cached key-value state. Watch: Best for teams already running model infrastructure. For small teams, start with simpler routing and prompt caching first. |
8.3/10 |
| n0-computer/iroh | 1.7k+ this week | Rust networking stack |
A modular networking layer built around dial keys rather than fragile IP assumptions. Interesting for local-first and peer-to-peer agent systems. Watch: Networking primitives are powerful, but security boundaries, identity, and key management matter more than demos. |
8.1/10 |
| penpot/penpot | 2.5k+ this week | Open-source design and code collaboration |
A mature open-source design tool matters because agents still need inspectable design systems, components, and handoff surfaces. Watch: Not an AI shortcut by itself. Its value is strongest when paired with a real design system and implementation workflow. |
8/10 |
| makeplane/plane | 1.5k+ this week | Open-source project management |
A modern Linear/Jira alternative for tasks, sprints, docs, and triage. Useful as a system-of-record for agent work. Watch: Agents need clean project data. If the backlog is messy, an open-source PM tool will not fix the workflow alone. |
7.9/10 |
| asgeirtj/system_prompts_leaks | 1.9k+ this week | System prompt collection |
Useful as education for prompt architecture, agent boundaries, tool policies, and product design patterns. Watch: Do not blindly copy leaked prompts. Treat this as study material, not a production policy source. |
7.6/10 |
| Kong/insomnia | 1.0k+ this week | Open-source API client |
A strong API client remains a core builder tool for testing REST, GraphQL, WebSockets, SSE, and gRPC. Watch: Keep environment variables, tokens, and workspaces clean before letting agents inspect API configs. |
7.5/10 |
| koala73/worldmonitor | 1.4k+ this week | Global intelligence dashboard |
Shows demand for situational awareness systems that aggregate news, geopolitics, infrastructure, and alerts. Watch: Intelligence dashboards can become noise machines. Source quality, filtering, and verification are the product. |
7.3/10 |
| iptv-org/iptv | 7.2k+ this week | Public IPTV channel collection |
Huge open index of publicly available channels. Useful for media discovery, monitoring, and player workflows. Watch: Respect local rights, content rules, and platform policies. Public does not always mean free for every use case. |
7.2/10 |
| freeCodeCamp/freeCodeCamp | 3.2k+ this week | Open-source education platform |
Still one of the best free learning ecosystems for programming, math, and computer science fundamentals. Watch: Foundational learning compounds, but it is not a direct agent stack component. |
7.1/10 |
| teslamate-org/teslamate | 431 this week | Self-hosted Tesla data logger |
Good example of durable self-hosted telemetry, dashboards, and personal data ownership. Watch: Vehicle and account integrations need care. Keep credentials secure and understand what data is being logged. |
6.9/10 |
| tursodatabase/turso | User-supplied trending pick | In-process SQLite-compatible database |
Interesting infrastructure for local-first apps and agent tools that need a small SQL layer without heavy server setup. Watch: Database choices should follow workload needs. Do not swap storage just because a repo is trending. |
7.7/10 |
Related Videos
I found a few related YouTube videos that help explain the direction behind the week: codebase memory, agent web reach, agentic video, and skill security. Treat these as supporting context and demos, not the factual source of truth for the repo claims.
codebase-memory-mcp related demo. Useful for seeing the code-memory idea in motion.
OpenMontage related demo. Shows the agentic video-production direction this repo is part of.
Agent-Reach related demo. Useful context for agent web access and search workflows.
SkillSpector related demo. Relevant to the new security layer around installable agent skills.
JQ AI SYSTEMS Top 10
If I had to choose only ten to study this week, I would choose based on workflow leverage, not raw popularity.
- DeusData/codebase-memory-mcp: Best overall builder pick: it reduces repeated codebase discovery and gives agents persistent code intelligence.
- Panniantong/Agent-Reach: Most useful research/access layer, as long as it is used with platform, account, and privacy boundaries.
- NVIDIA/SkillSpector: Most important safety pick because installable skills are becoming part of the agent supply chain.
- addyosmani/agent-skills: Best process pick for improving how coding agents plan, test, review, and ship.
- calesthio/OpenMontage: Best creative-system pick because video production is moving from manual editing into agent-readable pipelines.
- withastro/flue: Best sandbox direction for teams that want agents but do not want them loose in the main workspace.
- google-research/timesfm: Best applied-model pick for forecasting problems that already exist inside real businesses.
- LMCache/LMCache: Best model-infra pick for teams thinking seriously about cost and latency at scale.
- penpot/penpot: Best design workflow pick because agent-built products still need real design systems and handoff.
- makeplane/plane: Best operations pick because agent work needs tasks, docs, triage, and a clear system of record.
The top three are clear for me: codebase-memory-mcp, Agent-Reach, and SkillSpector. One helps the agent understand your codebase, one helps it reach the web, and one helps you inspect the skills you are about to trust. That is the new agent stack in miniature: memory, reach, and safety.
Patterns This Week
1. Codebase Context Is Becoming Infrastructure
codebase-memory-mcp, agent-skills, and Flue all point in the same direction: agents need durable context and controlled execution. Grepping every file again and again is not a strategy. Letting an agent run without a boundary is not a strategy either.
2. Web Access Is Still A Messy Superpower
Agent-Reach is attractive because it promises a single CLI for broad web reading and search. That is exactly what agents need. It is also exactly where teams can get into trouble with cookies, logins, terms of service, copyright, rate limits, and messy sources. Use it first for public, read-only research tasks.
3. Agent Skills Need Security Review
The rise of agent-skills, Claude skills, Codex plugins, and installable workflows makes SkillSpector timely. A skill is not just text. It can shape what an agent does, what it ignores, what tools it calls, and how it interprets success. Review it like a dependency.
4. Creative Workflows Are Becoming Programmable
OpenMontage is the most obvious creative repo this week, but the broader signal is bigger: video, editing, scripting, assets, and QA are becoming pipelines that agents can operate. The winner will not be the tool that generates the flashiest demo. It will be the workflow that keeps rights, brand, review, and repeatability under control.
5. The Normal Software Layer Still Matters
Penpot, Plane, Insomnia, iroh, Turso, and freeCodeCamp are a useful reminder: agents still need design systems, tasks, APIs, storage, networking, and human learning. The model does not remove the stack. It makes the stack more important.
Builder Checklist
- If your agent wastes time finding files: test codebase-memory-mcp.
- If your agent needs web research: test Agent-Reach on public read-only tasks first.
- If you install third-party skills: scan and review them with SkillSpector.
- If coding agents feel messy: try agent-skills before switching models.
- If you make repeatable videos: study OpenMontage, but keep editorial review and media rights in the loop.
- If you need forecasting: test TimesFM against your historical data before trusting any prediction.
- If you want an agent-ready product workflow: pair Penpot, Plane, and Insomnia with clear docs and review gates.
The CTA is simple: do not install the whole weekly trending page. Pick the one bottleneck that is slowing your agent workflow today: memory, reach, security, sandboxing, video, forecasting, design, PM, or APIs.
Sources
- GitHub Trending weekly, checked June 22, 2026.
- DeusData/codebase-memory-mcp
- Panniantong/Agent-Reach
- NVIDIA/SkillSpector
- addyosmani/agent-skills
- calesthio/OpenMontage
- withastro/flue
- google-research/timesfm
- LMCache/LMCache
- n0-computer/iroh
- penpot/penpot
- makeplane/plane
- asgeirtj/system_prompts_leaks
- Kong/insomnia
- koala73/worldmonitor
- iptv-org/iptv
- freeCodeCamp/freeCodeCamp
- teslamate-org/teslamate
- tursodatabase/turso
- YouTube: codebase-memory-mcp related demo
- YouTube: OpenMontage related demo
- YouTube: Agent-Reach related demo
- YouTube: SkillSpector related demo