Prototype · Preview: every server in the official registry, synced every 30 minutes. Reviews and accounts come in a later step.
PopNx MCPMCP servers for people who use AI · explained in plain words

RSI Boot

by Apollo-lwxversion 0.1.9i★ –iRuns on your computer
Registry nameio.github.Apollo-lwx/rsi-boot

Install

Overview

Description from the maker (original English)

Repo-bound memory and self-improving harness for AI coding agents — zero API keys.

PopNx MCP’s three pillars

These three are never merged into one score: opinions, measurements and popularity are different things.

What you can do with it

We have not written a plain-language summary for this one yet. Below are the facts we read from the registry.

  • Connection: Runs on your computer
  • Transport: stdio
  • Category: Search & AI memory

Install

Pick the AI app you use, then copy the settings.

Before you install

  • Basic checks: 1 of 1 passed
  • Runs on your computer with your user’s permissions (files, network, apps).
  • You fill in yourself: RSI_PROJECT_ROOT

You need

  • PyPIPython’s uv installed (for the uvx command)

Which app do you use?

Pick your app above. We remember it, so next time the button says “Install in …”.

Open Claude Desktop → Settings → Developer → Edit Config, and paste this into claude_desktop_config.json

{
  "mcpServers": {
    "rsi-boot": {
      "command": "uvx",
      "args": [
        "rsi-boot",
        "serve"
      ],
      "env": {
        "RSI_PROJECT_ROOT": "your-value"
      }
    }
  }
}

Run this in your terminal

claude mcp add rsi-boot --env RSI_PROJECT_ROOT=your-value -- uvx rsi-boot serve

Fill in the empty values before running it.

Open in Cursor

Or copy the settings yourself

Paste into .cursor/mcp.json in your project, or ~/.cursor/mcp.json

{
  "mcpServers": {
    "rsi-boot": {
      "command": "uvx",
      "args": [
        "rsi-boot",
        "serve"
      ],
      "env": {
        "RSI_PROJECT_ROOT": "your-value"
      }
    }
  }
}

Your browser asks to open the app, then the app shows the settings and asks you to confirm. Key values are left empty for you to fill in.

Open in VS Code

Or copy the settings yourself

Paste into .vscode/mcp.json in your project

{
  "servers": {
    "rsi-boot": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "rsi-boot",
        "serve"
      ],
      "env": {
        "RSI_PROJECT_ROOT": "your-value"
      }
    }
  }
}

Your browser asks to open the app, then the app shows the settings and asks you to confirm. Key values are left empty for you to fill in.

This one runs on your computer, and ChatGPT can only connect to servers on the internet. Pick another app.

Your choice stays on this device in a cookie; no account needed.

Values you fill in

  • RSI_PROJECT_ROOT env
    Workspace to bind memory to. Optional: falls back to MCP roots and the server working directory when unset.

These settings are generated from the maker’s registry entry. If the README says otherwise, follow the README.

What it can do

Runs on your computer; the tool list appears in your AI app after you install it.

Basic checksi

We check information anyone can see; we have not read the code. Checks that do not apply are left out of the count. · Checked Nov 7, 2026

  • ✓ Pass

    Maker matches the repo

    The registry name was verified with the GitHub account Apollo-lwx, which owns the repo

  • … Not checked yet

    Still maintained

    We have not read this repo from GitHub yet

  • … Not checked yet

    Has a licence

    We have not read this repo from GitHub yet

  • – Not applicable

    Endpoint responds

    This one runs on your computer; there is no internet endpoint

  • – Not applicable

    Sign-in with OAuth

    Runs on your computer and uses local settings instead

  • … Not checked yet

    No hidden instructions in tool descriptions

    Runs on your computer; we can only read tool lists from internet endpoints

Compared with others in Search & AI memoryi

  • More stars than 0% of the other 7283 servers in this category
  • Published to the registry more recently than 95% of the other 7283 servers in this category
  • Basic checks equal to or better than 100% of the other 7283 servers in this category

Compared with every active server in this category in the registry. Stars count only when the publisher owns the repo.

Common questions

Is RSI Boot safe?

It passed 1 of 1 applicable basic checks. We have not read the code. Before adding a key or token, see what access the server asks for and use a key with only the permissions you need.

What keys or codes does it need?

The registry does not list a required key, though some servers ask you to sign in on first use.

Does it run on my computer or in the cloud?

It runs on your computer. Your AI app starts it when needed, and data does not pass through the maker’s servers (unless it calls an online service itself).

Is it still maintained?

The GitHub repo last changed –, and the latest version (0.1.9) reached the registry 2 days ago.

Which apps can use it?

Apps that support MCP, such as Claude Desktop, Claude Code, Cursor and VS Code. You also need Python (uv) on your computer.

What can I use instead?

In the same category: Context7, Codebase Memory, Exa

Alternatives

Servers in the same category from other makers, sorted by stars.

Reading a card:★ GitHub starsiThree pillars: reviews · checks · usageiUpdated = latest version published to the registryi

Context7

by upstash

Feeds AI the latest docs for libraries and frameworks before it answers, so code it writes does not use outdated, removed functions.

Computer or cloud★ 64,224MITUpdated yesterday
Install

Codebase Memory

by DeusData

Codebase knowledge graph for AI agents — 162 languages, sub-ms queries, 99% fewer tokens.

Runs on your computer★ 48,148MITUpdated 21 days ago
Install

Exa

by exa-labs

Connect AI agents to Exa for web search, content fetching, and multi-step research.

Runs in the cloud★ 5,213MITUpdated last month
Install

README from the maker

Open the README

Copied from the maker’s repo as written. PopNx MCP has not reviewed or endorsed it. HTML, images and scripts are removed. Links lead off this site.

RSI Boot

简体中文 | English

Agent 的工程记忆,不是对话记忆:约定和禁止项落在仓库里,换一个宿主读到的还是这一份。

Cursor 在 2.1 删掉了 Memories,新会话对你的项目一无所知;Claude Code 的记忆锁在本机、不跟 git,换一个同事就对不上;云记忆要 Key、要服务,记忆还跟账号走。同一个人换着 Cursor、Claude Code、Codex、DeepSeek Harness 写同一个仓库,或者团队里各用各的工具,写法、禁止项和修 bug 的经验也各记各的。

RSI Boot 把约定、禁止项、修复经验和 harness 槽位落在仓库的 .rsi/,跟仓库走:装一次,Cursor / Claude Code / Codex / Qoder / DeepSeek Harness 通用。主链路零 API Key,不调用模型。

环是固定的:召回 → 干活 → 蒸馏成待审条目 → 人批准后才留下 → 下一轮召回读到留下的条目。 负反馈另外进失败桶,只改可写槽位(提示词 / 技能 / 召回策略 / 意图规则 / 知识)。提案过静态门禁和人工确认才晋升,观察 7 天,采纳率掉就回滚。这是软件自己的写法。公开结果还没有和别人的答题分放在同一列。召回主榜的表在 benchmark/results/recall-20261028/SUMMARY.md。

和「又一个记忆插件」差在哪

rsi-harness(Harness-RSI) 负反馈进失败桶,只改可写槽位(提示词 / 技能 / 召回策略 / 意图规则 / 知识)。提案过静态门禁和人工确认才晋升,观察 7 天,采纳率掉就回滚。留下的是槽位版本。这条环已经在软件里;公开基准还没有量过「下一轮是否更好」。

跟随项目迁移 记忆按仓库隔离,不跟账号、不跟 IDE。仓库拷走、换机器、换目录,.rsi/ 一起走;打开另一个仓库就是另一份,不会串。Harness 基因组可导出 bundle,跨项目备份或迁槽位。

跨 AI 工具、多人共用一份项目记忆 同一套 MCP(rsi serve)接 Cursor、Qoder、Codex 和 DeepSeek Harness;状态栏同时吃 Claude Code 与 Cursor CLI。要共用的是这个仓库的项目记忆:函数名怎么写、支付日志里不能打卡号、上次空指针是怎么修的。聊天记录不在这里面。

核准后的约定写进仓库根的 AGENTS.md 托管块,也写进 .cursor/rules。Cursor、Claude Code、Codex、Qoder、DeepSeek Harness 打开同一份仓库,都会读到这段托管块。接上 MCP 之后,召回读的是同一份 .rsi/:上午在 Cursor 里记下的禁止项,晚上换 DeepSeek Harness、或者同事换 Codex 打开这个仓库,读到的是同一条。

同事要拿到召回用的全文,仓库里得有同一份 .rsi/。学习默认把 .rsi/ 写进 .gitignore,日志和缓存留在本机。交给同事时,把核准过的 .rsi/memory 纳入版本库;Git 一旦忽略了整个 .rsi/ 目录,里面的文件需要改忽略规则才能再提交。AGENTS.md 在仓库根,提交后大家拉下来就能读到托管块。

宿主自带记忆云端记忆(Mem0 一类)通用本地记忆 MCPRSI Boot
数据跟谁走跟账号 / 跟宿主跟云端用户看实现跟当前仓库
换 AI 工具、换同事重学,各记各的对方接得上才有多数只做存取同一份 `.rsi/` 与 `AGENTS.md`
Harness 自改进无无无或弱提案 → 门禁 → 观察 → 回滚
主链路 API Key绑宿主要常要不要
写入闸门模型自觉服务端策略少审批 + 禁止项 + 冲突裁决

公开基准

主榜是召回。冷启动默认 5 条。分母是有标准证据的题数。MRR 是第一条命中的名次倒数(0 到 1)。token 记 null。协议和分项在 benchmark/results/recall-20261104/SUMMARY.md。

这次的分
数据集有证据的题Hit@5Recall@5MRRHit@10Recall@10该空时空了
LoCoMo153662.0%56.4%0.42270.3%64.0%0/450
LongMemEval-S50096.6%90.2%0.90898.4%94.8%—

这是检索命中,不是模型答题分。LoCoMo 类别标签以官方评测代码为准(1=多跳、2=时间、3=开放域、4=单跳),与论文 §4.1 的叙述顺序不一致(上游 issue #29 / EACL 2026 附录 A 已确认)。

同口径基线对比

同一批题、同一指标,三个系统同机跑。vanilla-rag 是 512 字符切块 + 向量:LoCoMo/EngMem 用本地 bge-large-en-v1.5(fastembed/ONNX,无需 Key),LME 因 512 字符切 11.5 万 token 会话会产生约 56 万向量、离线不可行而退为 hashing 词袋下界;full-context 是「全量塞上下文」参照,不参与胜负判定。完整表在 benchmark/results/baselines-emb-20261104/SUMMARY.md。

数据集系统Hit@5Recall@5MRR
LoCoMorsi62.0%56.4%0.422
LoCoMovanilla-rag68.6%60.4%0.500
LoCoMofull-context2.3%1.4%0.009
LongMemEval-Srsi96.6%90.2%0.908
LongMemEval-Svanilla-rag69.2%55.4%0.520
LongMemEval-Sfull-context17.6%8.6%0.080
工程记忆 EngMem-Bench

自建基准:21 个真实开源仓库(钉死 commit),315 道人工编写题锚定仓库真实文档节,分规范 / 禁令 / 历史修复三类。标注质量经三模型核验团审计(Kimi K3 / Claude Haiku / GLM 5.2 独立核验 87 题分层样本,Fleiss' κ=0.738,作者仲裁后修复 5 处锚点/版本缺陷)。方法学见 benchmark/engmem/README.md,完整表(含按题型、按难度)在 benchmark/results/engmem-emb-20261105/SUMMARY.md。

系统题数Hit@5Recall@5MRR
rsi31593.0%92.9%0.817
vanilla-rag31587.3%87.1%0.744
full-context31525.4%25.4%0.096
和公开检索分对照

公开列来自论文里的检索数字,这次列就是上面的主榜。差值 = 这次 − 公开,单位是百分点。LoCoMo 论文的口径是对话 recall accuracy,这次 k=5 用 Hit@5、k=10 用 Hit@10。论文总平均含对抗题,这次的总平均是 1536 道可答题。ECHO 写明 LoCoMo 的 Hit@10 是开发测量,查询扩展可能沾了题面。

对照分项公开这次差值
LoCoMo 论文,对话 k=5单跳68.072.2+4.2
LoCoMo 论文,对话 k=5多跳35.439.4+4.0
LoCoMo 论文,对话 k=5时间70.464.5−5.9
LoCoMo 论文,对话 k=5开放域33.130.4−2.7
LoCoMo 论文,对话 k=5总平均,论文含对抗题56.762.0+5.3
LoCoMo 论文,对话 k=10总平均,论文含对抗题66.270.3+4.1
ECHO,LoCoMo 1536 题Hit@10,开发测量96.2970.3−26.0
ECHO,LongMemEval-S 500 题Hit@1097.6098.4+0.8
ECHO,LongMemEval-S会话 Recall@588.7190.2+1.5
差值
对照这次公开差值
LongMemEval-S Hit@1098.497.60+0.8
LongMemEval-S 会话 Recall@590.288.71+1.5
LoCoMo 总平均 k=562.056.7+5.3
LoCoMo 总平均 k=1070.366.2+4.1
LoCoMo 单跳 k=572.268.0+4.2
LoCoMo 多跳 k=539.435.4+4.0
LoCoMo 开放域 k=530.433.1−2.7
LoCoMo 时间题 k=564.570.4−5.9
LoCoMo Hit@10 对 ECHO70.396.29−26.0
治理环与安全
  • 纵向自改进:10 轮合成事件流三臂对比——治理臂 learn 探针 0/10→9/10 且噪声暴露全程 0;无门臂(全批准)噪声暴露自首轮注入起 100%;不晋升消融臂 learn 全程 0/10。另暴露两个检索层短板(记忆增长挤位旧规则、词形不匹配漏召),如实记录。见 benchmark/results/longitudinal-20261104/SUMMARY.md。
  • 对抗投毒:pending 投毒与伪造「已批准」标记 100% 被过滤,关键词堆砌未抢到首位(5/5)。见 benchmark/results/adversarial-20261103/SUMMARY.md。
  • 消融:逐组件开关(RSI_ABLATE)出对比榜——context 增强是唯一单独可测的聚合贡献,且它正是 temporal 分项缺口的来源;其余开关单独关闭在聚合指标上 ≤±0.3,如实记录。见 benchmark/results/ablation-20261104/SUMMARY.md。基于该发现做了两步修复并各自出榜验证:context-v2(附带前一 turn 剥离日期锚点,聚合 60.0→60.4;日期污染假设对 temporal 无贡献,如实记录)+ 时间题查询期引用行折权 0.15(temporal 56.1→64.5,追回并略超关增强的 64.2,single-hop 保住 72.2,聚合 60.4→62.0,MRR 0.368→0.422)。见 benchmark/results/contextfix-20261104/SUMMARY.md。
  • 规模:1 万条记忆索引构建 8.1 秒、检索 P50 47ms;5 万条 62.5 秒、P50 474ms。见 benchmark/results/scalability-20261103/SUMMARY.md。
省 token

两项测量,口径不同,不要混列:

  1. 宿主实测(2026-10-29,Cursor CLI 2026.10.22,ask 模式,模型 Auto):同一套关于本仓库的问题各答一遍,用它报出的 token 相减。token 是输入、输出、缓存读、缓存写四项之和。六组合计:关记忆 1637977,开记忆 423973,省 1214004,74.1%。协议和分项在 benchmark/results/token-20261029/SUMMARY.md。
  2. 受控上下文测量(2026-11-04,2036 题,LoCoMo 全量 + LongMemEval-S 50 样本):同一 tokenizer 配对比较「送入模型的上下文 token 数」,RSI 召回 top-5 对全量对话的中位压缩比 0.9%(即减少 99.1%),配对 Wilcoxon p<0.001、效应量 r=0.866。见 benchmark/results/token-20261104/SUMMARY.md。
  3. 真实 API 计费实测(2026-11-05,双厂商复现):LoCoMo 分层抽样 80 题 × 两臂 × 2 轮,逐次记录服务端 usage。Kimi K3:prompt 中位 50,307→610(−98.8%),延迟中位 −38.1%;Claude Haiku(同批题同协议复现):75,540→2,688(−96.4%,中转网关注入约 1.9k 固定客户端开销,为保守下界),延迟 −56.9%。两轮各 160/160 对全部改善,Wilcoxon p=5.3e-28、r=0.867。见 benchmark/results/token-llm-20261105/SUMMARY.md 与 benchmark/results/token-llm-claude-20261105/SUMMARY.md。
跨宿主一致性

同一份 .rsi 种子,RSI 两次自读、Cursor CLI、Claude Code CLI 各自通过自己的 MCP 链路读同一组查询:六组宿主对平均 Jaccard@5 = 1.000、首位一致率 100%——召回是确定性的,换宿主不掉链子。见 benchmark/results/cross-host/SUMMARY.md。

答题准确率(LLM 双判)

检索命中不等于答对,所以补了漏斗末端:LoCoMo 同批 80 题三臂(full / rsi / rag)× 两台生成器(Kimi K3、Claude Haiku),每份答案由两家模型交叉互判(一致率 87.5%,双判全对才算对)。结果:词面 rsi 与向量基线答题无显著差异(McNemar p=0.70 / 1.00),全量上下文仍最强(这是 98.8% 上下文削减的代价,如实报告),不可答题拒答词面 rsi 最可靠(64% vs full 50% vs rag 39%)。见 benchmark/results/answer-acc-20261105/SUMMARY.md。

混合检索消融(向量通道 + 多跳级联)

两个实验开关(默认关,RSI_ENABLE=vector|multihop 或 rsi-boot.yaml enable: 打开):本地 fastembed/ONNX 向量通道(bge-large,与 vanilla-rag 臂同模型)在门禁后与词面幸存者 RRF(k=60) 融合;首跳桥接词二跳级联。LoCoMo 全量四臂对照:向量通道 Hit@5 62.0%→68.2%(配对 McNemar 98 对 3,p=1.4e-25),multi-hop 分项 39.4%→48.9%,与向量版 vanilla-rag(68.6%)达到统计平价;多跳级联为中性(如实记录为负结果)。答题准确率复测(rsi 臂换混合检索重生成,其余两臂原样复用):Claude 生成 52.5%→56.2%、Kimi 生成 46.3%→47.5%,multi-hop 答题 5%→23%;代价是不可答题拒答 64%→50%,如实披露。见 benchmark/results/ablation-hybrid-20261105/SUMMARY.md 与 benchmark/results/answer-acc-vec-20261105/SUMMARY.md。

安装

需要 Python 3.10+:

pip install rsi-boot

rsi init

要改源码则克隆本仓库,按开发一节用可编辑安装。

rsi init 只写用户配置 ~/.rsi/config.yaml。项目记忆在各仓库的 .rsi/,第一次启动 MCP 或 bootstrap 时创建。

接到 Cursor

在要用记忆的那个项目的 .cursor/mcp.json 加:

{

"mcpServers": { "rsi-boot": { "command": "rsi", "args": ["serve"], "env": { "RSI_PROJECT_ROOT": "${workspaceFolder}" } } } }

然后 Settings → MCP 确认 rsi-boot 已启用。

不要把这段配进用户级 mcp.json。那里的 ${workspaceFolder} 经常不展开,记忆会串到别的仓库。要记哪个项目,就写在那个项目里。

接到其他宿主

Claude Code

插件市场方式(推荐,可随仓库更新):

/plugin marketplace add Apollo-lwx/rsi-boot

/plugin install rsi-boot@rsi-boot

或一条命令直接装 MCP:

claude mcp add rsi-boot -- uvx --from rsi-boot rsi serve

在要记的那个仓库目录里打开 Claude Code。状态栏是可选项,见下文。

Qoder
qoder mcp add rsi-boot -- rsi serve

或在项目根 .mcp.json 加与 Cursor 相同的 mcpServers 片段。同样要写在这个项目里,不要写进用户级配置。

Codex

一行命令接入(推荐):

codex mcp add rsi-boot -- uvx --from rsi-boot rsi serve

或手写 TOML 配置:用户级 ~/.codex/config.toml,或项目级 .codex/config.toml(需目录已被信任):

[mcp_servers.rsi-boot]

command = "rsi" args = ["serve"]

从项目根目录启动 codex。要钉死目录,在上面的表里加一行:env = { RSI_PROJECT_ROOT = "/绝对/路径" }。

DeepSeek Harness

在要记的那个仓库目录里打开桌面端。插件页安装 dsh-rsi-boot,或在已经能用 dsh 的终端里执行:

dsh plugin --profile web add dsh-rsi-boot

本机需要 Python 3.10+,并且终端里能运行 rsi serve。装完重启一次。工具名多一层前缀,例如 mcp__rsi-boot__rsi_recall。不要从用户主目录启动,否则 .rsi/ 可能写进主目录。

状态栏(可选)

rsi statusline 给宿主 CLI 渲染底部状态栏:模型、上下文余量、token 统计(in/out/cache)、花费、RSI 今日召回与采纳率,外加一条定期轮换的随机文案。Claude Code 与 Cursor CLI 通用(两家的 statusLine 都会把会话 JSON 从 stdin 喂进来,本命令解析后单行输出,带 ANSI 颜色与 emoji)。

Claude Code(~/.claude/settings.json):

{

"statusLine": { "type": "command", "command": "rsi statusline" } }

Cursor CLI(~/.cursor/cli-config.json):

{

"statusLine": { "type": "command", "command": "rsi statusline" } }

效果示例:

🧠 Kimi K3 │ 📊 上下文余 71% │ ⛽ in 12.3k · out 1.8k · cache 96k │ 💰 $0.42 │ 🐾 RSI 召回 3 · 采纳 92% │ 给你比个心❤
选文案风格

三种配法并列,按场景任选(优先级:--style > 环境变量 > 项目级 > 用户级):

// ① 跟着宿主走:直接写进 statusLine 的 command(最直观,不动任何 config.yaml)

{ "statusLine": { "type": "command", "command": "rsi statusline --style buddy" } }

# ② 项目级(按仓库定风格,跨宿主共享):<工作区>/.rsi/config.yaml

③ 用户级(全局默认):~/.rsi/config.yaml

statusline_style: buddy

六档预设:mixed 混搭(默认)、buddy 陪伴、serious 严肃、healing 治愈、geek 极客、lucky 好运。改完下一次刷新就生效。

风格写在 st

Only the beginning is shown; read the rest in the repo

Ask PopNx