The “Open” wave: open-source rivals to Semrush, Composio, Claude Design and more, rated

Spend a week on GitHub's trending page in 2026 and a naming pattern jumps out: OpenSEO, OpenWorker, OpenWiki, OpenMontage, OpenCLI, OpenSpace. Each is pitched as the open-source version of a paid product, and many reached tens of thousands of stars within weeks. Stars measure attention, not readiness. We reviewed seventeen of these projects from their repos, docs and release notes. The useful question is the same each time: what does the project actually own, and what does it borrow from somebody else?

Personal agents and AI coworkers

OpenClaw (4.3) is the project the others get compared to. It is a self-hosted assistant that answers you in WhatsApp, Telegram, Slack, iMessage and more than 20 other chat apps. It has about 393,500 stars and had about 3.44 million npm downloads in a single week. It also runs its main tools without a sandbox by default and has had well over 100 security advisories, so read the security guide before you give it your shell.

OpenWorker (3.9), from Andrew Ng's account, takes the opposite approach. It is a desktop coworker whose approval steps and audit log are part of the design rather than added later. Its first specialist agents do security reviews. It is also a three-month-old beta with an unsigned Windows installer. OpenBot (3.8) from CopilotKit gives each bot its own sandboxed browser behind a gateway that refuses any action not explicitly allowed. It is a template you clone and adapt, though, not a product you install, and it relies on CopilotKit's paid memory service.

Marketing and research

OpenSEO (3.9), at about 25,900 stars, is pitched as an open-source Semrush. It owns no SEO data. Every keyword, rank and backlink number comes from DataForSEO, which you pay per query. For a small site that is a fair deal: hosted plans start at $10 a month, or you can self-host and pay DataForSEO directly. For anyone who needs Ahrefs-grade link data, it is not enough. We left out a similarly named project, open-seo-mcp-skills, because it is eight short Markdown files that point to a closed, hosted service.

OpenResearch (3.8) from alphaXiv turns Claude Code or Codex into a research agent with a paper search, a choice of compute back ends and an experiment log the agent checks before deciding what to try next. It sends analytics unless you opt out.

Trading

Two projects stand out among the trading agents. TradingAgents (4.0), at about 119,000 stars, has LLM analysts, bull and bear researchers and a risk team debate a stock until they reach a rating. Its recent releases tackle look-ahead bias, which makes it one of the few you can backtest honestly. Vibe-Trading (3.9), from the University of Hong Kong lab behind LightRAG, turns a plain-English strategy into a backtest across Chinese, US and crypto markets. It can place live orders, but only if you enable it and set the limits yourself. Neither is financial advice, and both cost you model tokens on every decision.

Building, docs and review

OpenWiki (4.2) is the most original idea in this list. LangChain's CLI writes a Markdown wiki of your codebase and ties every factual claim to the source lines it came from. When the code changes, only the pages whose evidence changed are regenerated. A run with no changes makes no model calls at all. Open Code Review (4.1) is Alibaba's release of the review tool it says it has used internally for two years. It decides in code which files need review rather than leaving that to the agent, and accepts lower recall in return for comments worth reading.

Open CoDesign (3.7) is a credible local Claude Design on your own model key, as long as you accept unsigned installers. OpenCLI (3.8) turns more than 100 websites into JSON commands and lets agents drive your logged-in Chrome. That makes it a credential: treat it with the same care as a password.

Infrastructure for agents

OpenConnector (3.8) is a self-hosted answer to Composio. It keeps SaaS credentials behind a gateway so the agent never sees a token, and it lists more than 1,000 providers. Its licence tightened after v1.9.4: you now need a commercial licence to resell it as a hosted service. Open Multi-Agent (3.6) has well-designed approvals and a run journal, but about 840 weekly npm installs against 7,000 stars suggests few people run it in production. OpenSpace (3.5) asks a good question, which agent skills actually work, but it lets agents rewrite their own skills by default and has had no commits since August.

Classroom and video

OpenMAIC (3.9) from Tsinghua turns a topic into an interactive lesson with AI teachers who talk and draw on a whiteboard. Its only protection by default is an access code. OpenMontage (3.7), at about 73,900 stars, turns a coding agent into a video production team. By its own prompt gallery, a short video costs between $0.15 and $3 on your own keys. It has one maintainer, no tagged releases and an AGPL licence.

How to read the stars

Three patterns show up across these reviews. First, many of these projects are a layer over someone else's data or model: DataForSEO, your coding agent, your LLM key. The open-source part is real, but the running cost lives elsewhere. Second, telemetry is often on by default even in projects that sell themselves on privacy, so check the environment variable before you deploy. Third, a large gap between stars and installs, or many stars and no releases, says a project has more fans than users. OpenClaw, OpenWiki and TradingAgents have both the attention and the track record. For most of the rest, pin a version and expect it to change under you.