Loopsy, a way for terminals and AI agents on different machines to talk

Establishes bidirectional communication channels between terminal instances running on different machines, allowing stdin/stdout/stderr streams to be transmitted over a network transport layer. Implements a client-server architecture where terminal I/O is serialized into messages and routed through a central coordination point, enabling remote command execution and output streaming without SSH or traditional shell tunneling.

Enables AI agents running on different machines to send commands and receive responses from each other through a standardized message protocol. Agents register themselves with a central broker, publish their capabilities, and can invoke remote agent functions by name with serialized arguments, implementing a publish-subscribe pattern for agent-to-agent RPC without requiring shared memory or direct network knowledge.

Captures terminal output (stdout/stderr) from a source machine and streams it in real-time to one or more destination terminals on different machines, maintaining character-level synchronization and preserving terminal control sequences. Uses a buffering strategy to handle network jitter and implements backpressure mechanisms to prevent buffer overflow when destination terminals lag behind source output rate.

Executes commands on remote machines while preserving or selectively overriding environment variables from the originating context. Implements environment serialization that captures the caller's environment state, transmits it to the remote machine, and merges it with the remote machine's base environment according to a conflict resolution policy (e.g., caller overrides, remote overrides, or merge with precedence rules).

Maintains a pool of persistent TCP/UDP connections between machines with automatic detection of connection failures and transparent reconnection without losing queued messages. Implements exponential backoff for reconnection attempts, message buffering during disconnection periods, and a configurable maximum buffer size to prevent unbounded memory growth. Reconnected clients automatically resync state with the server to ensure consistency.

Enables sequential execution of commands across multiple machines where the stdout of a command on one machine is piped as stdin to a command on another machine. Implements a pipeline orchestrator that manages data flow between machines, handles backpressure when downstream commands are slow, and provides error propagation so that failures in any stage halt the pipeline. Supports both linear pipelines (A→B→C) and fan-out patterns (A→B, A→C).

Provides a registry where AI agents can advertise their capabilities (available commands, input schemas, output schemas) to a central broker, and other agents can query the registry to discover available capabilities without prior knowledge. Implements a schema-based capability definition format (likely JSON Schema or Protocol Buffers) that describes input parameters, output types, and execution constraints. Discovery queries support filtering by capability name, tags, or required input types.

Captures the complete state of a terminal session (command history, environment variables, working directory, output) and serializes it to a portable format that can be replayed on another machine or at a later time. Implements a session log format that includes timestamps for each I/O event, enabling deterministic replay with original timing or accelerated playback. Supports filtering and searching within session logs to extract specific commands or output segments.