Context Crumb

MCP办公协作

提供方 未标注

办公协作 stdio · 本地进程 未标注 社区 / 第三方 ClaudeCursorVS CodeWindsurf
查看接入方式

将长文件、提示输入和MCP目录描述压缩为更密集的上下文,供LLM代理使用,同时保留有用信号。

一句话选型结论

按资料字段整理

Context Crumb 属于办公协作类 MCP 服务器,stdio · 本地进程,鉴权方式未标注,可直接接入 Claude、Cursor、VS Code 等,有可直接执行的安装命令。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类办公协作 传输方式stdio · 本地进程 鉴权未标注 部署方式本地部署
GitHub 星标 未标注 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

办公协作
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

生产力

安装与配置

stdio · 本地进程 未标注
客户端在本机拉起子进程,数据不出本机
npx skills add Yuchen20/Context-Crumb

这条命令只是入口,具体启动参数请按官方文档补齐,不要直接照搬到别的服务器。

  1. 1
    准备运行环境

    stdio 服务在本地起进程,需要先装好 Node.js 等运行时,再按安装命令拉取依赖。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
日程与消息

查日程、发消息、约会议。

表单与审批

填表单、走审批流。

团队协作

同步任务、汇总进展。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 有现成的安装命令,不用自己找入口,照着跑就能起服务。
  • 已适配 Claude、Cursor、VS Code,换客户端不用重新研究配置。

先想清楚的地方

  • stdio 方式要求本地装好运行时(如 Node.js),换机器要重新装一遍。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在办公协作场景里让模型直接调用外部能力的人。
  • 已经在用 Claude、Cursor 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 完全不想在本地装环境的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
Context Crumb当前 未标注 stdio · 本地进程 本地部署 有 未标注
CowAgent 未标注 stdio · 本地进程 本地部署 有 2026-04-11
Karakeep MCP 未标注 未标注 未标注 无 2026-05-17
agentscope 未标注 stdio · 本地进程 本地部署 有 2026-04-11
note-gen 未标注 stdio · 本地进程 本地部署 有 2026-04-11
Everywhere 未标注 stdio · 本地进程 本地部署 有 2026-04-12

服务器档案

办公协作
字段来自项目仓库与公开索引,中文说明由本站标注
提供方未标注上游标注的维护方 未标注
分类办公协作本站按上游分类展示
连接方式stdio · 本地进程客户端在本机拉起子进程,数据不出本机
鉴权方式未标注远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · 未标注本地安装时用到的包标识与运行时
部署类型本地部署上游给出的部署形态
最近核验未标注上游最后核验时间

详细介绍

8.6 KB
来自项目仓库与上游目录的原文,未做删改

ContextCrumb

Shake the crumbs out of bloated context.

=3.10">

Before / After -
Quickstart -
Playground -
Install -
CLI -
Agent + MCP -
Model


LLM context gets messy fast: notes, logs, issue threads, docs, research dumps, and tool descriptions all pile up until the useful signal is buried under filler.

ContextCrumb is a token-level compressor for LLM and agent workflows. It looks at text word by word and removes low-signal tokens while keeping the surviving text in the original order.

That is the idea behind the name: the context is still there, but the loose crumbs are shaken off before they reach your model. Less bloat in the prompt. More room for the parts that matter. Less wasted usage when Codex, Claude Code, or another agent processes long files repeatedly.

No install needed. Paste text, compare the kept context, and see what gets shaken off.

Before / After

ContextCrumb is not a summarizer. It does not rewrite your document into a new explanation. It keeps the source sequence and deletes expendable words. This example uses target_keep_ratio=0.72.

Original```text
Agents spend context on notes, logs, tickets, docs, and tool descriptions. Those files contain useful facts, but they also carry filler phrases and repeated wording. ContextCrumb compresses the text before it reaches the model. It keeps the original order, removes low-value tokens, and leaves a shorter version with the names, actions, constraints, and sequence still intact.


**Compressed**```text
Agents spend context notes, logs, tickets, docs tool descriptions. Those files useful facts, carry filler phrases repeated wording. ContextCrumb compresses text before reaches model. keeps original order, removes low-value tokens, leaves shorter version names, actions, constraints sequence intact.

Same order. Less padding. More room for the next file. On prose-heavy agent inputs, ContextCrumb often saves around 30-70% of the context depending on how aggressively you compress and how much filler is in the source.

Metric Original Compressed Saved
Model tokens 72 52 20 tokens
Token budget 100% 72% 28% fewer input tokens

What that feels like over a month

Assume your agent reads 8k-token notes, logs, tickets, research dumps, or docs before answering. This helps with API token bills, but also with subscription-based coding agents where heavy context reads can burn through usage faster.

Workflow Files read / day Context saved / month API cost avoided at $5 / 1M input tokens Subscription usage feel
Solo agent helper 20 ~1.4M-3.4M tokens ~$7-$17 Fewer bulky reads in Codex or Claude Code
Busy project workspace 200 ~14M-34M tokens ~$72-$168 More room for actual reasoning and edits
Agent-heavy team or eval loop 2,000 ~144M-336M tokens ~$720-$1,680 Less usage spent processing padded files

The bigger win is usually not only the bill. It is keeping long-running agents from filling their context, turns, and subscription usage with words they did not need to carry in the first place.

Quickstart (30-second setup)

Teach your agent a small habit: compress the bloat before it enters context. ContextCrumb is meant to sit in the background as a skill, stepping in whenever a long note, doc, issue thread, research dump, or log would otherwise flood the context window and eat into your Codex or Claude Code usage.

  1. Add the skill.```bash
    npx skills add Yuchen20/Context-Crumb

2. Select the agent you want to install it on.

The skill tells your agent when to compress text, how to preserve the useful sequence, when supported code can be loaded with comment/docstring compression, and when exact raw text is required for configs, direct quotes, or exact edits.

3. Use ContextCrumb to compress long files instead of dropping the whole thing into context.```text
Use ContextCrumb to compress this long project note before you work from it.
  1. Voila: every long note, log, ticket, research dump, or doc enters context already trimmed, saving tokens and preserving more of your agent subscription for the work that matters.

Why ContextCrumb?

Use case What changes
Agent file loading Compress long notes, docs, research dumps, and logs before they hit the context window.
Prompt pipelines Shrink natural-language inputs without hand-writing summarizers.
MCP catalogs Compress verbose tool/resource descriptions while preserving names and schemas.
Local workflows Run ONNX inference by default, with cached model files after first download.
Subscription-aware agents Spend less Codex or Claude Code usage on repeatedly loading padded prose.
Inspection and tuning Use diff and inspect to see what was kept, deleted, and saved.

Best fit: docs, notes, issue threads, logs, research context, other natural-language files, and supported source files where only comments/docstrings should be shortened. For exact code edits or exact comments, read the raw source.

Install```bash
pip install contextcrumb


Optional extras:```bash
pip install "contextcrumb[mcp]"
pip install "contextcrumb[serve]"
pip install "contextcrumb[torch]"

ContextCrumb uses the ONNX backend by default, so normal users do not need PyTorch or Transformers installed. Model files are cached locally after the first download.

CLI

The main agent-friendly command is load:```bash
contextcrumb load notes.txt


It prints only compressed text by default, which makes it easy for agents, hooks, shell scripts, and prompt pipelines to capture stdout and move on. For subscription tools like Codex or Claude Code, that means fewer bulky file reads before the agent gets to the useful part.

Useful commands:```bash
contextcrumb load notes.txt --json
contextcrumb load notes.txt --receipt
contextcrumb config set compression.content_mode auto
contextcrumb diff notes.txt
contextcrumb inspect notes.txt
contextcrumb stats

--receipt leaves compressed text on stdout and writes a compact savings receipt
to stderr. ContextCrumb uses compression.content_mode = "auto" by default:
prose files are compressed normally, while supported code files use a
code-aware path that preserves executable source exactly and compresses only
comments/docstrings. Unsupported syntax-sensitive files such as diffs, configs,
lockfiles, SQL, and .env files are still refused unless you pass --force;
forced output is only for exploratory reading, not exact edits or copy-paste
commands.

Persistent defaults live in user config and can be changed from the CLI:```bash
contextcrumb config show
contextcrumb config set compression.content_mode code-comments
contextcrumb config set code.comment_target_keep_ratio 0.55
contextcrumb config unset compression.content_mode


Supported file modes:

| Mode | Behavior |
| --- | --- |
| `auto` | Prose uses normal compression; supported code uses `code-comments`. |
| `prose` | Treat the whole input as natural language. |
| `code-comments` | Preserve executable code exactly and compress only comments/docstrings. |
| `raw` | Return the file unchanged with stats. |
| `refuse` | Reject file compression. |

Initial code-aware languages: Python, JavaScript, TypeScript, JSX, TSX, Go, and
Rust.

`diff` marks deleted tokens like this:```text
kept words [-deleted words-] kept words

Agent + MCP

ContextCrumb includes an optional MCP stdio adapter for agent clients that can run Python tools through uvx.```bash
pip install "contextcrumb[mcp]"


Published-package MCP config:```json
{
  "mcpServers": {
    "contextcrumb": {
      "command": "uvx",
      "args": [
        "--from",
        "contextcrumb[mcp]",
        "contextcrumb-mcp"
      ]
    }
  }
}

The MCP server exposes:```text
compress_text
compress_file


ContextCrumb also ships `contextcrumb-shrink`, an MCP proxy that compresses verbose catalog descriptions before an agent sees them while forwarding tool names, schemas, calls, results, and resource contents unchanged. This is useful when an agent client repeatedly spends context and subscription usage just looking at long tool descriptions.

Model

Model weights and a hosted demo are public on Hugging Face:

- Model: ymao20/contextcrumb-32m
- Playground: contextcrumb-32m-demo

## Roadmap

Planned for later:

- Public docs for advanced compression modes and service deployment.
- JavaScript or TypeScript client.
- Hosted API experiments.
- npm publishing.

## Development```powershell
uv pip install --python .\.venv\Scripts\python.exe -e ".[dev,mcp]"
.\.venv\Scripts\python.exe -m pytest
.\.venv\Scripts\python.exe -m build

Release notes are tracked in CHANGELOG.md.

License

MIT. See LICENSE.

常见问题

8 问
答案由本站依据条目字段整理,不引用上游原文
Context Crumb 怎么安装?

在客户端按下面的安装命令接入:npx skills add Yuchen20/Context-Crumb。装完在客户端配置里引用同一个包或端点。

Context Crumb 怎么在 Claude、Cursor、VS Code 里配置?

它的传输方式是stdio · 本地进程,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

Context Crumb 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

Context Crumb 需要 API Key 吗?

资料没有标注鉴权方式,是否需要密钥取决于服务本身,接入前按官方文档确认。

Context Crumb 支持哪些客户端?

资料标注支持 Claude、Cursor、VS Code、Windsurf、Codex、Zed;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

Context Crumb 能提供哪些能力?

资料显示能力标签包括 生产力。具体工具清单以官方文档为准。

Context Crumb 需要什么运行环境?

属于本地进程,需要在本机装好运行环境。

Context Crumb 是本地还是远程?

它是本地进程(stdio),由客户端在本机拉起。传输方式为stdio · 本地进程。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 未标注。