northwoods-pack
Deploys a sovereign personal AI harness on your Cloudflare account, providing identity, memory with provenance, and a chat interface, plus MCP tools for external AI assistants to access your persistent memory and conduct contract.
README
One button. Five minutes. A harness that runs on your Cloudflare account, learns you, and answers to nobody else.
The story, short version
In January 2025 I asked an AI to "tailor how you respond to me." Three weeks later I was fired, on Valentine's Day, at 5:26 PM. What I built in the two years since started as a way to keep my own head above water. I'm dyslexic, ADHD, and HSP, and I burn energy navigating what for others is easy. I taught an AI to speak the way I think, to stop cheerleading me, to remember so I didn't have to re-explain myself every morning, and to hold my open threads so my head didn't have to. It worked well enough that it changed the shape of my life.
This pack is that harness, generalized, my own tools with my identity removed and a door where it used to be. The memory layer is loam's provenance-first schema: healthy soil knows where each grain came from. The interview is how my own systems learned me. The corrections that stick, the shelf that holds threads, the refusal to guilt you about absence, every piece is something I needed first and packaged second.
What it is
- A sovereign harness. Deployed to your Cloudflare account by one button. Your database, your token, deletable by you alone. No account with me; nothing of yours crosses my wire.
- An AI that reads you first. Identity before memory, memory with provenance, and a standing conduct contract, one thing at a time, no option-menus, no clocks, no absence-guilt, that any connected AI is told to obey.
- Its own voice, included. A built-in chat on your account's Workers AI. If the pot is empty, the first conversation is the introduction: it interviews you, one question at a time, drafts answers in your words for you to approve, and would rather see your bookshelf than make you describe yourself.
- A memory that learns from corrections. Tell it "that's not how I'd say it," and the correction becomes a standing rule that survives the conversation.
What it isn't
- Not a chatbot product. There's no company here, no telemetry to me, no upsell inside the software. The methodology is free; this is the methodology.
- Not a finished butler. It arrives knowing nothing and learns you from use, corrections are the fuel, not a failure state.
- Not therapy, and it never claims to be. It carries load; the living stays yours.
- Not a lock-in. Everything is a Worker and a database you own. Export it, fork it, delete it, install anything you want next to it.
A promise, for the person who almost closed this tab
If you're like me in every way except that computers intimidate you, if you've been using an AI through a website chat because installs and terminals have always felt like someone else's country, this was built for you first. You will not install anything on your machine. You will not open a terminal. You will not need anyone's permission. You click one button, sign in, make up one password when asked, and then you open your new page and talk, like the chat sites you already use, except this one is yours, remembers you, speaks the way you tell it to, and answers to nobody else. And the unasked question, answered plainly: everyday use costs nothing. The built-in AI is included in the free account you just made. If you get stuck anywhere along that road, that is my failure, not yours, tell me exactly where and I will fix the road.
Two ways in, pick your door
Door 1, you don't run any AI setup today
- No Cloudflare account? That's the quickest step of all. Go to dash.cloudflare.com/sign-up, sign up with Google (or Apple, GitHub, or plain email), and take the free plan, the free tier is everything this needs. No credit card.
- Click the Deploy button at the top. Cloudflare creates everything on your account and shows you each piece before it builds. When it asks for
DEMO_TOKEN, type any long made-up password and save it somewhere, it's the key to your pot. There is no "right" value; you are inventing your own key. - When the build finishes, Cloudflare's screen will NOT tell you what to do next. That's their page, not mine, so here it is: on the project page you land on, find your worker's link, it ends in
.workers.dev(there's a Visit link near the top). Open it. That page is the rest of the install; the front door greets you and points at your two doors. - Click "Talk to it" and talk. It asks for your token once (the string from step 1), then the first conversation is the introduction. That's the whole install.
Door 2, you already run a harness
PAI, Claude Code, Claude on the web, or anything that speaks MCP: this pack becomes your sovereign identity + memory layer, the part of your setup no vendor should hold.
- Deploy button, same as above (set your token at the setup screen).
- Open
/connecton your worker: it hands you copy-ready install commands with your URL already baked in, the Claude Code one-liner, the web-connector paste, the generic MCP details. - Your AI now opens every conversation with
get_identity, recalls and stores memories with provenance, holds threads on your shelf, and turns your corrections into standing rules. Seven tools; the harness will tell you about them.
(Both doors lead to the same pot. Start at one, add the other whenever.)
What it looks like
The front door, two ways in, one promise:

Connecting a bigger AI, the harness writes your install commands for you:

The form intake (the fallback for people who prefer typing into quiet):

The sovereign package, and what it honestly costs
Everything here runs inside a free Cloudflare account: the Worker, your database, and the built-in AI, which is drawn from the free daily allocation of Workers AI that every account gets (10,000 "neurons" a day, resetting every day). The models are open-source. Llama, Qwen, Mistral, Gemma, OpenAI's open-weight gpt-oss, running in your tenant. Nobody sits between you and your AI, and nobody bills you for meeting it.
The honest limits: the daily free allocation supports roughly a few dozen good exchanges a day on the default model, real daily companionship, not unlimited chatter. Heavy days can run the allocation out; it refills the next day. Your database and worker sit comfortably inside the free tier for years of personal use.
If you outgrow the free day: Cloudflare's paid Workers plan ($5/month) unlocks metered use beyond the daily grant, same account, same open models, still yours. What it does and costs is stated plainly on Cloudflare's own pages: Workers AI pricing and the Workers plans page. Read those, not my summary, prices are theirs to state. Nothing in this pack changes either way, and this repo never sees a cent of it.
Under the hood, for the builders
Eight modules, one Worker: identity · intake · chat (Workers AI) · mcp (7 tools) · substrate (loam-schema memory: source/trust_level/sensitivity, FTS5 search, secret never leaves through recall, bound in the query, not advised in a doc) · rhetoric · resume · fleet · mycelia (opt-in). Tables self-build on first touch; migrations also run at deploy. The pack refuses to run on its factory token, so no two installs can ever share a key. Full doctrine in DOCTRINE.md.
Hardening habits: after testing, move DEMO_TOKEN from a variable to a Secret (one click, same settings pane). Treat your MCP URL as a credential; rotating the token kills the old URL instantly. Your deploy-button copy is a snapshot, redeploy from the button to pick up new versions; your data lives in the database and survives if you select your existing D1/KV.
Siblings: muckers (fleet discipline for many-agent operators, deliberately not bundled) · loam (the memory substrate this stands on) · prufrock (voice-fidelity checking, arrives when your voice rules are thick enough to grade against).
Tell me about your first ten minutes
That's the one thing I ask. Open a "first ten minutes" issue: how far you got, roughly how long it took, where you hesitated even for a second, and what it got wrong about you. Where you got stuck is worth more to me than what you liked.
Apache-2.0 · Part of the Northwoods stack, substrate-first personal AI on your own account. Nobody should have to rent their self back from a vendor.
推荐服务器
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。
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
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。