Nagi
A prototype phone built around an agent-centric interaction paradigm. Its home screen has no app grid; a locally run LLM-driven agent handles intent understanding, memory, and execution, while Android applications are reduced to callable low-level capabilities.
Timeline (last 90 days)
A developer demonstrated a prototype phone project called Nagi, whose core premise is that the home screen is not a grid of app icons but an Agent that is executing tasks on the user's behalf, with interaction primarily by voice or text.
Nagi downgrades regular Android apps to underlying capabilities that can be invoked but no longer occupy a home screen entry point; apps can still be installed and used.
Nagi uses a local LLM to handle intent recognition, memory, and context processing; unless the user actively connects to a cloud model, context data does not leave the phone.
Nagi adopts a model-agnostic design, and users can connect OpenAI, Anthropic, local models, or use the default configuration.
Nagi's execution path prioritizes MCP / API calls, with directly driving native Android apps as the fallback, but that path is described by the developer as rather rough.
On the hardware side, Nagi keeps only one physical button: press and hold to speak, tap to record a note.
Experiences in Nagi are reusable multi-step flows, such as calling a car to go home, planning a trip, or clearing the inbox and drafting replies; the user can generate a new flow by describing it once in natural language, without a builder or code.
All Facts (7)
A developer demonstrated a prototype phone project called Nagi, whose core premise is that the home screen is not a grid of app icons but an Agent that is executing tasks on the user's behalf, with interaction primarily by voice or text.
50%UnverifiedNagi downgrades regular Android apps to underlying capabilities that can be invoked but no longer occupy a home screen entry point; apps can still be installed and used.
50%UnverifiedNagi uses a local LLM to handle intent recognition, memory, and context processing; unless the user actively connects to a cloud model, context data does not leave the phone.
50%UnverifiedNagi adopts a model-agnostic design, and users can connect OpenAI, Anthropic, local models, or use the default configuration.
50%UnverifiedNagi's execution path prioritizes MCP / API calls, with directly driving native Android apps as the fallback, but that path is described by the developer as rather rough.
50%UnverifiedOn the hardware side, Nagi keeps only one physical button: press and hold to speak, tap to record a note.
50%UnverifiedExperiences in Nagi are reusable multi-step flows, such as calling a car to go home, planning a trip, or clearing the inbox and drafting replies; the user can generate a new flow by describing it once in natural language, without a builder or code.
50%