Honcho Server

Maintains persistent, evolving models of individual user psychology and behavior patterns across conversation sessions using a server-side state store. Honcho tracks user preferences, communication styles, emotional patterns, and interaction history to build longitudinal profiles that inform subsequent interactions. This enables the AI to adapt responses based on accumulated knowledge of each user rather than treating each session as stateless.

Manages concurrent multi-user conversation sessions with isolated state contexts and participant role tracking through MCP resource endpoints. Honcho coordinates interactions between multiple participants (users, agents, moderators) within bounded session contexts, maintaining separate psychology models and interaction histories per participant while enabling cross-participant reasoning and coordination.

Enables agents to reason about how user psychology, preferences, and goals have evolved over time, identifying trends, inflection points, and long-term patterns in user behavior. Honcho maintains timestamped psychology models and enables agents to query historical snapshots, compare user state across time periods, and identify significant changes or patterns.

Enables long-running AI reasoning tasks to execute asynchronously outside the request-response cycle, with results stored and retrievable via MCP resource endpoints. Honcho decouples expensive reasoning operations (multi-step planning, user psychology inference, cross-participant analysis) from immediate response requirements, allowing agents to perform deep analysis in background tasks and reference results in subsequent interactions.

Enables AI agents to reason about other agents' and users' mental states, beliefs, goals, and likely actions using structured theory-of-mind models exposed via MCP. Agents can query what they believe about other participants' knowledge, preferences, and intentions, then condition their responses on these models. This implements a form of recursive reasoning where agents model not just user behavior but user understanding of the agent.

Implements the Model Context Protocol (MCP) server specification to expose Honcho capabilities as standardized resources, tools, and prompts that any MCP-compatible client can invoke. Honcho acts as an MCP server, defining resource schemas for user psychology models, session state, and reasoning results, and implementing the MCP transport layer (stdio, SSE, or custom) for client communication.

Automatically extracts and infers user preferences, values, and communication styles from conversation history using LLM-based analysis and stores them as queryable preference profiles. Honcho parses user messages to identify stated preferences, infers implicit preferences from behavior patterns, and maintains a structured preference model that agents can query to personalize responses without explicit user configuration.

Manages LLM context windows by selecting the most relevant conversation history and user psychology information to include in prompts, using importance scoring and psychology-aware ranking. Rather than simple recency-based truncation, Honcho ranks conversation turns by relevance to current user psychology state and agent goals, ensuring that psychologically-significant interactions are retained even if older.

Builds and maintains models of agent behavior patterns, decision-making styles, and reasoning approaches based on historical agent actions and outputs. Honcho tracks how agents respond to similar situations, what reasoning patterns they use, and how their behavior evolves, enabling prediction of agent actions in novel scenarios and detection of anomalous behavior.

Streams interaction events (messages, actions, state changes) to analytics backends in real-time, enabling live monitoring and post-hoc analysis of user-agent interactions. Honcho emits structured events for each interaction turn, psychology model update, and reasoning step, allowing downstream systems to build dashboards, run analytics, and audit agent behavior.

Defines and manages function schemas that agents can call to take actions, with built-in support for multi-agent function coordination and result aggregation. Honcho allows agents to declare available functions via JSON schemas, call functions with type-safe arguments, and coordinate function calls across multiple agents (e.g., one agent calls a function that another agent must handle).