driftglass

driftglass

MCP server that maintains cited, current answers to standing research questions by tracking chosen sources, consolidating repeated coverage, and providing evidence-based briefs with change signals.

Category
访问服务器

README

Driftglass

Keep a current answer to questions that outlive the news cycle. Driftglass follows the sources you choose, consolidates repeated coverage, and maintains a cited answer with the causes, strongest alternative case, and signals that would change it.

Driftglass keeping an AI infrastructure answer connected to its supporting evidence

Try the frozen, read-only live example or watch a standing question move from new sources to a saved answer in the quick walkthrough.

Product tour

Today Mission workspace
Three standing questions in the revised Driftglass Today view The AI infrastructure Mission workspace with its current files
Public answer Architecture
A cited public answer prepared by Driftglass Driftglass architecture from sources to a current answer

Quick start

1. Install on Cloudflare

Activate R2 in the target Cloudflare account, then install Git and Node.js 22 or newer:

git clone https://github.com/anotb/driftglass.git
cd driftglass
npm install
npx wrangler login
cp .deploy-secrets.example .deploy-secrets
# Replace the placeholder with the output of: openssl rand -hex 32
npm run check:deploy
npm run deploy:first

This creates the Worker resources, applies D1 migrations, and sets 30-day expiry for disposable raw/ captures. Direct public-page reading works on Workers Free. Browser Rendering, Email Routing, AI Search, the Companion, and Workers Paid are optional. See Deployment for costs, staging, and later upgrades.

2. Run the experimental local preview

The source-checkout preview requires Git and Node.js 24.4 or newer:

git clone https://github.com/anotb/driftglass.git
cd driftglass
npm install
npm run selfhost:build
node dist/selfhost/driftglass-selfhost.mjs init --data-dir "$PWD/.local-driftglass"
node dist/selfhost/driftglass-selfhost.mjs serve --data-dir "$PWD/.local-driftglass"

This starts a loopback-only instance with local SQLite and filesystem storage. It is not yet a packaged desktop or server release. See Portable Runtime for backups, restore, and current limits.

3. Connect ChatGPT or another reasoning model

After Driftglass is running, open Reasoning in the dashboard.

  • ChatGPT: choose Connect ChatGPT, complete the connection, then select Driftglass from Tools in a new Work chat. The optional @Driftglass plugin gives steadier Today and Mission routing.
  • Claude: add the compact remote MCP, or use the generated Agent Skill and Markdown/JSON context.
  • Other agents: connect the read-only MCP at /mcp. Machine-readable discovery lives at /llms.txt, /llms-full.txt, /.well-known/mcp.json, /.well-known/driftglass.json, and /openapi.json on the running instance.

Add the separate Allow updates connection only when the agent should save answers or suggestions for your review.

WebMCP remains experimental. It needs Chrome 149 or newer, an origin-trial token for the deployed origin or chrome://flags/#enable-webmcp-testing locally, and an open, unlocked Driftglass dashboard tab. The compact remote MCP works without WebMCP.

First useful session

  1. Open the dashboard with the owner secret and run the setup check.
  2. Keep the Free usage profile and install one featured Intelligence Pack.
  3. Collect once. Driftglass groups repeats into Stories and matches them to the Pack's Research Missions.
  4. Open Today and choose a Mission with a question you care about.
  5. Connect your reasoning model and ask for the current answer, what changed, and what would change it again.
  6. Save the cited answer if it is worth carrying into the next update.

Start there. The Companion, AI Search, Email Routing, and Computer Power Mode can wait until a Mission needs them.

What Driftglass keeps for each question

A Research Mission holds:

  • the question and current answer
  • the sources and developing Stories behind it
  • chronology, disagreements, and missing information
  • the next event or measurement worth watching
  • earlier answers, decisions, forecasts, and outcomes

Missions can cover energy flows through Hormuz, AI entering the experimental science cycle, a market thesis, a policy change, or a technical dependency. They are standing questions, not feed categories.

chosen sources → collected items → developing Stories → Research Mission
                                                           ↓
Today ← cited current answer ← chosen reasoning model ← prepared context
  ↓                                                        ↑
what changed → next event → scheduled refresh → connected memory

Quiet days do not need a briefing.

Sources and answers

Ten articles repeating one announcement are still one line of reporting. Driftglass groups sources by publisher, repository, package, author, community, and declared family, then marks their relationship to a Story:

origin · independent · same-family · update · echo

Briefs favor original reporting and independent support while retaining useful updates. They include the Mission question, source links, relevant memory, disagreements, open questions, and the requested answer shape. Each brief reports what it included and omitted; a connected model can use search, fetch, or a larger prepared context when it needs more of the source history.

The returned answer keeps its citations, model label, confidence, and source-state hash. A second model can answer from the same source state for comparison. Suggested memory changes stay pending until approved.

Driftglass does not require an OpenAI, Anthropic, or xAI model API key. It uses compatible subscription clients through MCP or portable context files.

What Driftglass includes

  • Today: a finite briefing of changed Missions and Stories.
  • Research Missions: standing questions with a current answer, next event, and history.
  • Connected memory: links among sources, claims, findings, questions, decisions, and forecasts.
  • Intelligence Packs: reusable sources, Missions, starter memory, standards, and scheduled research for a domain.
  • Mission Computers: one working directory per Mission for its current source snapshot, notes, results, and exports.
  • Sharing: cited web pages, downloadable copies, and credential-free Packs for following the same public question.
  • Budget Governor: keeps collection, rendering, storage, and scheduled work within the selected usage profile.

The default MCP is compact and read-only. Its operations counterpart uses an independently derived capability and handles saved results, outcomes, Pack changes, workspace notes, and approved memory changes.

Sources

The cloud core supports Hacker News, Lobsters, Bluesky, arXiv, GitHub Releases and Activity, npm, PyPI, public webpages, Page Feeds, Email Workers intake, and manual capture. OpenAlex uses an optional account key.

The optional Companion adds signed-in sources and mirrors Mission workspaces to macOS, Windows, or Linux. Browser sessions stay on that machine. Public collection and Mission state continue without it.

Architecture in one minute

Layer Role
Workers API, dashboard, routing, MCP, and static assets
D1 Stories, Missions, memory, decisions, budgets, and run state
R2 source bodies, reasoning inputs, results, exports, and previews
Queues and Workflows bounded ingestion, recovery, Mission refresh, and research routines
Direct HTTP; optional Kitesurf and Chromium public-page reading and richer rendered-page fallbacks
Cloudflare Computer one filesystem per Mission
Drop portable answers and Intelligence Packs

Cloudflare is the full hosted profile in 0.9. The local SQLite profile is an experimental source-checkout preview. Deployment creates a dedicated per-environment OAuth KV namespace. AI Search stays off until enabled, and retained Workers traces are disabled.

Architecture details: Product Boundary · Runtime Fabric · Reasoning Interfaces · Memory Graph · Mission Computer · Intelligence Packs · Budget Governor

Docs

Acknowledgements

Driftglass builds on and learns from open-source projects including OpenCLI and Cloudflare Computer. See Acknowledgements for the full list and each project’s role.

License

MIT. Integrated upstream tools remain separately installed under their own licenses.

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选