SELF-HOSTED · HERMES AGENT · CLAUDE · PI
X-Agent-Webui 是为超长 Agent 对话打磨的自托管工作台。几十万 token 之后，它仍然记得你说过的每一句话。
X-Agent-Webui is a self-hosted workspace hardened for agent conversations that run to hundreds of thousands of tokens — and still remember every word you said.
真实界面截图，会话内容为演示数据。Actual interface; conversation content is demo data.
01 · MEMORY-FIRST COMPACTION
普通的上下文压缩会把长对话"总结掉"，你早先定下的约束也一起消失。X-Agent 把用户意图当成最高优先级的上下文来保护。
Ordinary context compression summarises a long conversation away — and the constraints you set early on vanish with it. X-Agent treats your intent as the context that must survive.
摘要失败、超时或返回错误提示时一律保留原文——历史绝不会被一句错误替换。压缩进度在对话里实时可见，完成后上下文数字立即刷新。
A failed, timed-out or error-shaped summary is discarded and the original kept — history is never replaced by an error string. Progress shows in the chat, and the context meter refreshes the moment it finishes.
一段约 45 万 token 的对话，生成摘要所需时间：
Summarising a conversation of about 450K tokens:
发送给辅助模型前先脱敏；辅助模型不可用时自动退回由 Claude 自己写。
Content is redacted before it reaches the auxiliary model; if that model is unavailable, Claude writes the summary itself.
02 · BUILT FOR THE LONG RUN
每条消息带稳定身份，贯穿发送、排队、多端广播、断线恢复与落库。刷新、重连、重启后不重复不丢失，不会一直卡在「工作中」，中止前已流出的输出也会保存。
A stable identity follows every message through sending, queueing, multi-tab broadcast, reconnect recovery and storage. Refreshes, reconnects and restarts neither duplicate nor drop anything, runs never hang on "working", and output streamed before an abort is kept.
Hermes Agent、Claude、Pi 共用同一界面与会话体验。推理强度可按对话单独设置（低到 Max）并真实生效，上下文用量显示准确，对话自动起标题。
Hermes Agent, Claude and Pi share one interface and one session model. Per-conversation reasoning effort (Low to Max) really takes effect, context meters are accurate, and conversations title themselves.
记忆锁：不解锁，Agent 就无法写入你的 MEMORY.md / USER.md，切换即时生效。整理好的笔记可一键同步给 Claude。
The memory lock keeps agents from writing your MEMORY.md / USER.md until you unlock it, effective immediately. Curated notes sync to Claude in one click.
界面冻结，不随上游频繁改版；每次升级只更新 Hermes Agent、Claude Code、Pi 三个 Agent 层。群聊、工作流等冗余功能已移除，专注对话本身。
The interface is frozen instead of churning with upstream redesigns; upgrades only refresh the Hermes Agent, Claude Code and Pi layers. Group chat, workflows and other extras are gone — the focus is the conversation.
03 · PERFORMANCE
在约 800 个会话的真实使用环境中实测，改造前后对比：
Measured in real use with about 800 sessions, before and after:
0.8 s
改造前Before ~3 s
14 MB
改造前Before 126 MB
0.2 s
改造前Before ~2 s
~59 fps
改造前Before 12–20 fps
0.3–0.5 s
改造前Before 1.3 s
0.1–0.4 s
改造前Before 1.5–2 s
数字来自本发行版的真实使用环境，实际表现取决于硬件、模型与服务商。
Figures come from real use of this distribution; results depend on hardware, models and providers.
04 · INSTALL
$ curl -fsSL https://raw.githubusercontent.com/gimee/x-agent-webui/main/install.sh | bash
默认账号 admin / 123456。请先在本机登录并立即修改用户名和密码，再考虑开放到公网。
Default credentials are admin / 123456. Sign in locally and change both immediately before exposing the service beyond loopback.
05 · LINEAGE & LICENSE
X-Agent-Webui 是独立维护的非官方社区衍生版本，基于 EKKOLearnAI/hermes-studio（Hermes Studio）v0.7.18，不代表上游官方，也未获其背书。
X-Agent-Webui is an independently maintained, unofficial community derivative of EKKOLearnAI/hermes-studio (Hermes Studio) v0.7.18. It is not affiliated with or endorsed by the upstream project.
代码遵循 BSL 1.1 及其非商业使用授权（Additional Use Grant）：可自托管与非商业使用，商业使用需向原作者取得授权；该许可将于 2029-05-10 转为 Apache-2.0。底层 Hermes Agent 由 Nous Research 以 MIT 许可发布。
The code is under BSL 1.1 with its non-commercial Additional Use Grant: self-hosting and non-commercial use are allowed, commercial use requires the original licensor's permission, and the license converts to Apache-2.0 on 2029-05-10. The underlying Hermes Agent is released by Nous Research under MIT.
