Just weeks after OpenAI unveiled Dots, its vision for persistent AI agents that operate in dedicated cloud environments, users have flooded forums with complaints. Tasks fail midway. Connections drop without warning. Browsers inside the virtual machines crash with cryptic error codes. The promise of always-on coworkers that handle complex workflows independently has met the messy reality of early deployment.

Developers didn’t wait for fixes. Within days of the launch, open-source projects surfaced on GitHub that replicate much of what Dots offers. Some run locally on a Mac using personal API keys. Others let users swap in models from DeepSeek, Qwen or Kimi through services like OpenRouter. The speed of imitation surprised even seasoned observers in the AI community.

The Register first chronicled this burst of activity. Tawkit, the company behind the CopilotKit SDK, dropped OpenDots last week. “OpenDots is completely open source and gives you the same thing as a template you host yourself,” the firm stated in its announcement. The project supports any agent framework via AG-UI, works with OpenAI-compatible models, and allows full customization of every component.

But. The original Dots from OpenAI carry a steep $200 monthly price tag for full access. They demand reliance on OpenAI’s infrastructure. Open-source versions sidestep both constraints. One popular fork, simply called Open Dot, runs directly on a user’s Mac. It supports local models or routes through OpenRouter. Credentials stay protected in the system keychain. The agent never sees passwords.

Reports of instability continue to mount. On October 7, OpenAI acknowledged multiple incidents affecting Dots. Thread creation failed in Codex and Work modes. Performance degraded across related services. Browser sessions inside Dots crashed repeatedly with Error code 9 and CDP recovery timeouts, according to user posts on the OpenAI Developer Community. Voice handoffs between spoken requests and task coordinators produced errors like “spoken context could not be confirmed.” One demo during OpenAI’s Dev Day earlier faltered when a Dot simply stopped responding mid-presentation.

These hiccups haven’t slowed the copycats. X posts from the past week show dozens of repositories gaining stars rapidly. One thread listed six viable alternatives, from Hermes Agent by Nous Research to lightweight Python implementations with long-term memory. “OpenAI Dots costs $200 a month and GitHub already has it for free,” one developer wrote. The post detailed projects that add features the official version lacks, such as cron-style scheduling for subagents or stricter approval gates before any file changes.

The underlying technology draws from OpenAI’s earlier o1 series. Those models introduced hidden chain-of-thought reasoning trained through reinforcement learning. They spend compute at inference time to think through problems step by step before answering. Dots appear to build on that foundation but add persistence, tool use, and a dedicated virtual computer. Yet replicating the full reasoning power has proven difficult for open-source teams.

Separate efforts target the reasoning layer itself. Repositories like o1-imitator collect papers and projects that approximate o1 behavior using techniques such as Monte Carlo Tree Search, self-refinement, or process supervision. One listed project, Open-O1, attempts to recreate the test-time compute scaling that gives o1 its edge on hard benchmarks. Performance still trails the original in many cases. But progress arrives fast. A recent DeepSeek reasoning model reportedly matches o1 on formal math benchmarks at a fraction of the compute cost, according to discussions on X.

OpenAI itself has pushed boundaries in related areas. On the same day The Register article appeared, the company published 722 mathematical manuscripts generated by an unreleased internal model. The Next Web covered the release, which groups results into 372 families across multiple domains. The model tackled roughly 4,000 open problems. Each result consumed about three hours of thinking compute on average. Some proofs were formally verified in Lean. Others await human review. Mathematicians advising OpenAI called the dump a starting point rather than a finished product.

The math release highlights what persistent agents like Dots could eventually achieve. If an agent can run uninterrupted for hours, applying structured reasoning across tools and documents, it might tackle research-level work. Yet current technical difficulties suggest that vision remains distant. Browser crashes break web automation. Voice integration fails to hand off context. Persistent state sometimes vanishes between sessions.

So the open-source community has taken a pragmatic path. Instead of duplicating the full cloud agent with all its reliability headaches, projects focus on self-hosted templates. Users bring their own models. They define approval workflows that pause before risky actions like git pushes or financial transactions. Memory layers from projects like Mem0 persist preferences across runs. Tool integrations via Composio connect to more than 1,000 applications without exposing secrets.

One X thread from October 7 cataloged over a dozen such components. Persistent browser control. Sandboxed code execution. Calendar and email access with human oversight. The pieces exist today. Assembling them no longer requires OpenAI’s infrastructure or pricing. “The interesting part of Dots is not the avatar,” the post noted. “It is the combination of persistent state, a computer that stays available, tools connected to real work and a hard boundary around actions that still require you.”

Competition has intensified. Chinese firms linked to Moonshot AI reportedly attempted to extract hidden reasoning traces from OpenAI models earlier this year, according to Decrypt. The campaign involved thousands of coordinated prompts designed to make the internal chain of thought visible for distillation into other systems. OpenAI disrupted the effort. The incident underscores how valuable the reasoning process has become.

Industry watchers expect this pattern to continue. Every time OpenAI ships a new capability, forks and approximations appear within days. The o1 models faced similar scrutiny when released in 2024. Developers quickly built systems that simulated long reasoning traces even if they couldn’t match the trained reinforcement learning behind them. Dots have accelerated that cycle because they expose agentic behavior in a more tangible form. Users can see the agent browsing, typing, and pausing for approval. That visibility invites replication.

Challenges remain for the imitators. Local models often lack the raw capability of frontier systems. Running persistent agents consumes significant resources on personal hardware. Integration bugs surface when swapping between providers. Still, the momentum feels unmistakable. Four thousand stars on the OpenDots repository in a single week tell their own story.

OpenAI has responded to some criticism by promising improvements. Outages on October 6 and 7 were resolved within hours. Updated desktop clients fixed thread creation problems. Yet users continue posting screenshots of failed browser sessions and unresponsive voice commands. The gap between marketing claims of “always-on” agents and delivered reliability has created an opening that open source is filling with speed and transparency.

What emerges next may not look exactly like Dots. It could be a federation of specialized agents, each with narrow responsibilities, running across mixed local and cloud backends. Or a fully offline system that trades some intelligence for complete privacy and control. The technical difficulties OpenAI faces have not discouraged builders. They have instead inspired a wave of experimentation that may ultimately define how agentic AI reaches the workplace.