Atlasdepth

MCP文档知识

提供方 未标注

文档知识 未标注 会话登录 社区 / 第三方 Claude DesktopClaudeCursorVS Code

面向 AI 智能体的本地优先屏幕记忆。借助证据锚点搜索你自己的屏幕历史记录。

一句话选型结论

按资料字段整理

Atlasdepth 属于文档知识类 MCP 服务器,传输方式未标注,会话登录,可直接接入 Claude Desktop、Claude、Cursor 等,没有现成安装命令,需按官方文档接入。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。

分类文档知识 传输方式未标注 鉴权会话登录 部署方式未标注
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

文档知识
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

记忆

安装与配置

未标注 会话登录
上游未给出传输方式
  1. 1
    准备运行环境

    按官方文档确认运行环境(常见为 Node.js 或 Python),装好后再执行安装命令。

  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
文档检索问答

在你的文档库里按问题找答案。

知识库读写

把新资料写进知识库,随用随取。

内容整理

把零散资料归纳成结构化笔记。

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

  • 已适配 Claude Desktop、Claude、Cursor,换客户端不用重新研究配置。

先想清楚的地方

  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 使用前要准备好会话登录对应的凭据,并按最小权限配置。
  • 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

  • 需要在文档知识场景里让模型直接调用外部能力的人。
  • 已经在用 Claude Desktop、Claude 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

  • 不方便管理密钥或第三方授权的人。
  • 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
Atlasdepth当前 未标注 未标注 未标注 无 2026-11-20
dify 未标注 stdio · 本地进程 本地部署 有 2026-04-11
open-webui 未标注 stdio · 本地进程 本地部署 有 2026-04-11
JeecgBoot 未标注 stdio · 本地进程 本地部署 有 2026-04-11
Cherry Studio 未标注 未标注 未标注 无 2026-05-17
mempalace 未标注 stdio · 本地进程 本地部署 有 2026-04-11

服务器档案

文档知识
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
分类文档知识本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式会话登录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
部署类型未标注上游给出的部署形态
最近核验2026-11-20上游最后核验时间

详细介绍

8.7 KB
来自项目仓库与上游目录的原文,未做删改

AtlasDepth · Memory Layer for AI agents (MCP over stdio)

One line: give the AI agent on your own machine a way to ask
"where is that quotation I saw last week?" — and get back an answer with its
provenance
, instead of handing over your whole screen history.

This repository publishes the memory adapter only: a single-file,
standard-library-only MCP server that reads a local AtlasDepth instance
(screen_memory.db) read-only.


Why this exists (the problem it solves)

Agents are good at writing things and bad at remembering what actually happened.
AtlasDepth records what you saw on screen locally (event-driven, ~1 frame per change),
OCR-indexes it, and keeps the original image next to the text. This adapter lets
an agent retrieve anchored memories — time, app, window title, a text snippet,
and the path to the source image — so the agent can answer and show where it came from.

That last part is the point: a citation you can click back to.
A model that is "usually right" is everywhere; a record that can prove itself is not.


The four boundaries (they are why the compliance story still holds)

# Boundary What it means in practice
1 stdio — no port, no daemon The agent spawns our subprocess and talks JSON-RPC over stdin/stdout. Nothing listens on any port, so no other program can fetch your memory. (This is deliberate: an HTTP endpoint would destroy this property.)
2 Deny by default, per-session consent Without explicit consent, every tool except memory_scope (which returns no memory content at all) is refused — and the refusal tells the agent how to ask the human to authorize. Consent is time-limited and revocable with one command.
3 Minimal scope, private never leaves Only the current project + last N minutes, at most 20 items per call. Rows with visibility=private are filtered in code, per row.
4 Anchors, not dumps Default payload is {id, time, app, window title, text snippet (160 chars), image path, confidence}. Full OCR text only on an explicit second call (memory_evidence) — which also keeps the agent's token bill sane.

An anchor is exactly this (real field names — write your integration against these):```json
{
"id": 1,
"time": "2026-10-02 11:47:14",
"app": "winword.exe",
"title": "报价单-示例.docx",
"text": "报价单 甲方:示例客户 乙方:我方 数量 3 台 单价 500 元 合计 1500 元",
"file": "C:\path\to\instance\screenshot\演示_报价单.jpg",
"conf": "高",
"truncated": false
}


* `text` is capped at 160 chars and flagged with `truncated`; ask `memory_evidence` for the full text by `id`.
* `conf` is a **human-readable band** (`高` / `中` / `低` — the app's own wording), not a float.
* `file` is the path to the **original screenshot** on that machine — that is what makes the answer citable.

Every call — **including refused ones** — is appended to a local audit log
(time / which agent / which tool / scope / how many items returned).

---

## Tools

| Tool | Consent needed | Returns |
|---|---|---|
| `memory_scope` | **no** | whether consent is active, minutes left, scope description. No memory content, not even the project name. |
| `memory_recent` | yes | anchored memories from the current project within the last N minutes — args `{minutes, limit}` (default 30, max 20) |
| `memory_search` | yes | keyword search within the current project, same anchor shape — args `{query, limit}` (`query` is required) |
| `memory_evidence` | yes | full OCR text + image path for one anchor — args `{anchor_id}` (the `id` from either tool above); returns `{"item": {…same fields, `text` no longer truncated…}}` |

---

## Quick start

### Try it in 2 minutes (no AtlasDepth install needed)

The adapter reads a local AtlasDepth instance — but you do not need one to see how it behaves.
Build a **fictional** demo instance (it contains "示例客户 / 样例发票 / 00000001" only), and the
script prints the exact registration snippet, the authorize/revoke commands, three questions to ask
your agent, and **what counts as success**:```bash
python demo/造演示实例.py --to ./demo_instance

That instance has a record from another project, a private row, an over-long text and a
row outside the 30-minute window — so you can watch all four boundaries hold in one run.
Delete the folder when you are done; it touches nothing else.

With your own AtlasDepth instance

The app's folder is named 深图 — that is its Chinese name. Use whatever your folder is actually called.```bash

1) check status (deny by default)

python mcp_memory_server.py --instance "/path/to/your/深图" --status

2) authorize for 30 minutes — this is the human step

python mcp_memory_server.py --instance "/path/to/your/深图" --grant 30

3) after you are done, cut it off; and see who asked what

python mcp_memory_server.py --instance "/path/to/your/深图" --revoke
python mcp_memory_server.py --instance "/path/to/your/深图" --audit 20


Register it in any MCP-capable client (stdio server), e.g. Claude Desktop's
`claude_desktop_config.json`:```json
{
  "mcpServers": {
    "atlasdepth-memory": {
      "command": "python",
      "args": ["/absolute/path/to/mcp_memory_server.py",
               "--instance", "/path/to/your/深图"]
    }
  }
}

Registering does not grant access — you still have to run --grant.
Recommendation: --revoke when you are done. Do not leave it authorized permanently.

The instance directory is where the AtlasDepth app keeps settings.json and
screen_memory.db (the app itself is Windows-only for now). You can also set
DSH_INSTANCE instead of passing --instance.


How to get the full app (this repository is only half of it)

This repository publishes the agent-facing half: the adapter, plus a fictional demo instance so
you can watch the boundaries hold without installing anything.

The other half — the Windows desktop app that actually records your screen locally, OCR-indexes it
and keeps the original images next to the text — is not publicly distributed yet. It is in a
small two-week trial right now, on purpose: the thing worth testing is not "another recorder", it is
what people actually manage to retrieve, and that needs a handful of real users rather than a
download button.

If you want to try the whole thing:

  1. open an issue titled Try request (there is a template for it), and
  2. say in one or two lines what you would want it to help you remember (e.g. "quotations and
    invoices I only ever see once", "meeting notes I can't find again").

That is the honest state of it. It also means "how many people asked" is a number we can actually
observe — which matters more to this project than looking finished.


Verify it yourself

The boundaries above are not marketing copy — they are enforced in code, and the
test suite drives a real subprocess over JSON-RPC:bash python tests/selftest_mcp_memory.py

It builds a throwaway instance containing a private row, a row from another
project
, and an over-long text, then asserts, among others:

  • nothing but memory_scope works before consent;
  • visibility=private rows never appear in any tool;
  • other projects' rows do not leak;
  • item count and snippet length are capped; full text needs the second call;
  • the audit log records refused calls too;
  • after --revoke, calls are refused again immediately;
  • the source contains no network primitives (no socket, no HTTP server).

License

This adapter is open source — Apache License 2.0 (see LICENSE). You may use, modify and
redistribute it, including commercially; keep the license notice and the attribution below
(Apache-2.0 §6 does not grant trademark rights, so please do not present a modified build as the
official one).

Copyright © 2026 SHF

The AtlasDepth app itself is not open source. It is free to use (personal and internal business use;
see the one-page license notice shipped with the app), it is not published as source, and it is
handed out through the trial request route above — please do not redistribute the app.


Not done yet (read this before assuming)

  • No packaged build / no download: the app half is Windows-only and is handed out through the
    trial request route above, not as a public binary. This adapter is a source script — running it
    needs Python 3 on your machine (that is also why the demo instance exists).
  • No GUI for consent: authorizing/revoking is a command, not a switch in the app.
  • Read-only: agents cannot write into the memory through this adapter. Writes
    only happen through the AtlasDepth app itself, driven by the human.

常见问题

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

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

Atlasdepth 怎么在 Claude Desktop、Claude、Cursor 里配置?

它的传输方式未标注,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。

Atlasdepth 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。

Atlasdepth 需要 API Key 吗?

资料标注需要会话登录认证,使用前准备好对应凭据,并按最小权限配置。

Atlasdepth 支持哪些客户端?

资料标注支持 Claude Desktop、Claude、Cursor、VS Code、Cline、Codex;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。

Atlasdepth 能提供哪些能力?

资料显示能力标签包括 记忆。具体工具清单以官方文档为准。

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 2026-11-20。