clone.is

What It Is

Clone is a “persistent user model” tool for AI Agents. Its goal is to predict what a user is likely to type in response to a given Agent prompt, so the Agent does not have to repeatedly interrupt the human when confidence is high enough. It is not a general-purpose chatbot; it is more like an intent-prediction and memory layer between the user and the Agent.

Core Capabilities

The product uses a three-layer architecture: Recording → Memory → Prediction. Recording captures user-generated events, including computer-usage frames, terminal interactions, Agent prompts and responses, integrated Webhooks, and more. Memory distills those records into UserProfile, SemanticMemory, EpisodicMemory, and RawMemory. Prediction then assembles context based on memory, calls Anthropic to generate top-K candidate replies, and returns calibrated confidence, raw confidence, and an auto/escalated decision. Clone places particular emphasis on confidence calibration: the system performs daily calibration based on whether users accept, reject, or edit its suggestions.

API, Integrations, and Deployment

Clone’s API design is relatively restrained. The core endpoint is POST /api/predictions/predict/. The response includes predicted_response, ranked candidates, confidence, and status. It is also MCP-native: the MCP Server exposes seven tools, including predict_next_prompt, predict_continuation, submit_feedback, and record_agent_prompt, and can connect to clients such as Claude Code, Cursor, and Claude Desktop. For deployment, the documentation explicitly supports self-hosting, with docker-compose able to start Postgres, the Django API, the MCP Server, and the Web container.

Pricing and Chinese Support

The collected text does not disclose any free tier, trial policy, plan pricing, or payment methods, so its commercial cost cannot be assessed. Chinese-language support is also not explained. Since its predictions depend on user data and Anthropic calls, Chinese performance would need to be tested in practice and cannot be confirmed from the text alone.

Pros, Cons, and Limitations

Its strengths are clear positioning, suitability for Agent automation scenarios, confidence-scored outputs that make threshold setting easier, and support for MCP and self-hosting, which makes it developer-friendly. The limitations are also obvious: cold-start prediction confidence is only in the 0.3–0.5 range and improves only after real user history accumulates; privacy, encryption, data retention, and compliance details are not disclosed; and it depends on Anthropic, with unclear model choices and cost structure.

Who It’s For and Access from China

Clone is better suited to AI Agent developers, heavy users of coding assistants, and teams that need to build a personal memory layer or self-hosted prediction service. It is less suitable for ordinary non-technical users. Access from mainland China is not discussed in the text; if it depends on Anthropic, actual use may also be affected by network and payment conditions. Alternative directions include OpenAI Assistants/Responses API, LangGraph, Mem0, Zep, Letta, Dify, and Coze.