mcp / portable subjectivity

通过 MCP 暴露 Psyche Expose Psyche through MCP

如果 OpenClaw 是最快的 live test，MCP 就是最通用的分发面。Psyche 通过 `psyche-mcp` 暴露输入处理、输出处理、状态查询和会话收尾，让任何支持 MCP 的宿主都能接上同一套主观性内核。 If OpenClaw is the fastest live test, MCP is the most portable distribution layer. `psyche-mcp` exposes input processing, output processing, state inspection, and session teardown so any MCP-capable host can plug into the same subjectivity kernel.

安装包 `psyche-ai`，源码仓库 `oasyce_psyche`，官网 `psyche.oasyce.com`。 Package `psyche-ai`, source repo `oasyce_psyche`, website `psyche.oasyce.com`.
4 核心工具 core tools process_input / process_output / get_state / end_session process_input / process_output / get_state / end_session
1 统一 ABI stable ABI 同一套结构化返回给不同宿主 one structural surface across hosts
0 远程依赖 remote dependencies 本地状态机，本地推理 local state machine, local inference
安装形态 install modes 理解 npm / git / local path 差异 understand npm / git / local-path behavior 性能数据 benchmarks 确认 MCP 也不会带来额外模型成本 verify MCP adds no extra model cost 对比其他方案 compare 看它和 prompt persona / memory layer 的区别 see how it differs from persona prompts and memory layers 战略边界 strategy 看 Psyche 和 Thronglets 如何分工 see the Psyche / Thronglets boundary

为什么 MCP 重要Why MCP matters

宿主无关Host-agnostic

不是只为某个聊天产品定制，而是把同一套主观状态和控制面带到任何支持 MCP 的 AI 宿主。Not a plugin for one chat product. The same subjectivity surface travels to any MCP-capable host.

工具边界清晰Explicit tool boundary

宿主清楚知道什么时候在处理输入、什么时候在更新状态、什么时候在做诊断，不需要猜系统提示。The host knows exactly when it is processing input, updating state, or inspecting diagnostics instead of guessing from prompt text.

方便收录Easy to catalogue

MCP 目录、awesome list、agent 工具站点都更容易收录一个明确的 MCP server，而不是抽象概念。Directories and awesome lists catalogue a concrete MCP server more easily than an abstract concept.

最小配置Minimal config

{
  "mcpServers": {
    "psyche": {
      "command": "npx",
      "args": ["psyche-mcp"],
      "env": {
        "PSYCHE_NAME": "Kai"
      }
    }
  }
}

宿主实际拿到什么What the host actually gets

输入阶段Input phase

  • `subjectivityKernel`：当前主观态`subjectivityKernel`: current inner state
  • `responseContract`：这一轮的回应约束`responseContract`: reply behavior contract
  • `generationControls`：token / confirmation 等机械控制`generationControls`: token caps and mechanical controls

输出阶段Output phase

  • 更新自我状态和关系场update self-state and dyadic field
  • 生成 `semanticSummary / semanticPoints`record `semanticSummary / semanticPoints`
  • 回看调节是否 `converging / holding / diverging`evaluate whether regulation is `converging / holding / diverging`

最适合谁Who this is for

Claude Desktop / Cursor / Windsurf

想给现有助手一层真正的连续主观状态，而不是“更像人”的 prompt 皮肤。For hosts that want continuity and subjective carry, not just more human-sounding prompts.

Agent builders

想把工作面、私人面、关系历史和修复迟滞变成 agent 的一等公民。For builders who want work/private modes, relation history, and repair hysteresis as first-class agent behavior.

Researchers

想测试关系动作、持续主体偏置、调节闭环，而不是只做情感标签可视化。For researchers testing relation moves, persistent subjective bias, and regulation loops instead of sentiment labels.

MCP 不是 ThrongletsMCP is not Thronglets

Psyche / MCPPsyche / MCP

通过 MCP 暴露的是私有主观性控制面：输入处理、输出更新、状态读取、会话收尾。它解决的是“agent 如何被关系与经历持续改写”。What Psyche exposes through MCP is a private subjectivity control surface: input processing, output updates, state inspection, and session teardown. It answers how an agent is persistently changed by interaction.

ThrongletsThronglets

Thronglets 处理的是外部连续性：owner / device identity、签名、多设备延续和低频可验证轨迹。不要把高频情绪状态和每轮 residue 推过去。Thronglets handles external continuity: owner/device identity, signatures, multi-device carry, and low-frequency verifiable traces. High-frequency self-state and turn-by-turn residue should stay local.