Your AI Agent Forgets Everything. This Project Gives It a Portable Cloud Brain

You've probably noticed the pattern by now: you set up an AI agent in one tool, teach it things, wire up credentials, and then switch to a different MCP host—and suddenly it's a blank slate. Every host keeps its own little silo of context, and nothing travels with you. Kody is an attempt to fix that by giving your agents a shared cloud of their own—memory, keys, code, and automations—that stays portable across every MCP host.

What It Does

Kody is, in the project's own words, "your agents' cloud." It's a system that holds the state your AI agent accumulates—packages, jobs, credentials, memories, and related data—and exposes it through OAuth-protected MCP endpoints. Because it's built on the Model Context Protocol, any MCP host can talk to it, which is the whole point: your agent's context isn't locked to a single application.

Architecturally, it's a Cloudflare Workers project through and through. There's a TanStack Start website, Worker-based request routing, package runtime plumbing, and the MCP endpoints themselves. The repo is an Nx monorepo with shared modules, an origin app worker, sibling workers for platform, runtime, and jobs, plus mock Workers for testing. The data layer is Cloudflare D1, and the project follows several conventions from the epicflare starter template.

One design choice worth noting up front: Kody deliberately favors a compact MCP surface. Rather than dumping a huge static catalog of tools on the model, it leans on powerful search and Code Mode execute flows. That's a meaningful decision about how agents should interact with tools, and it shapes the whole product.

Why It's Cool

• Portability is the real feature. The pitch isn't just "store agent memory"—it's that the memory, keys, and code live somewhere host-agnostic. If you've ever rebuilt the same context in three different agent tools, you know why that matters.

• It handles teams, not just individuals. Accounts support orgs, teams, and grants, and each account gets isolated packages, jobs, credentials, and memories. The README is explicit that no account is privileged at runtime—tests may seed deterministic local accounts, but there's no special admin backdoor baked into the system. That's a clean security posture for a multi-tenant tool.

• The compact MCP surface is a smart bet. A smaller tool catalog with strong search and an execute mode means less context bloat and more flexible behavior. It's the kind of choice that looks boring on a feature list but pays off in practice.

• OAuth-protected endpoints are the default, not an afterthought. Given that this thing stores credentials, that's the right call.

• The stack is boring in the best way. Cloudflare Workers, D1, TanStack Start, Nx, npm. If you're already in that ecosystem, there's very little new infrastructure to learn.

How to Try It

Getting the dev environment running is about as short as it gets:

npm install
npm run dev

The dev server starts at localhost:3742 by default—though the CLI will pick the next free port if that one's taken and print the resolved URL. Wrangler handles the local Cloudflare Workers runtime and the D1 database automatically, so you don't need to provision anything by hand.

A few pointers before you dive in:

1. Read docs/contributing/getting-started.md for the full setup paths and what to expect.
2. If you're contributing (or an agent working in the repo), start with AGENTS.md for repo-specific guidance.
3. If you're trying to figure out what the project is actually for, docs/contributing/project-intent.md is the place to begin.
4. Production topology details live in docs/contributing/architecture/index.md.

If you'd rather scaffold a fresh project from the underlying epicflare template instead of using Kody directly, the README points to npx create-epicflare.

You can find the repo at github.com/kody-hq/kody, and there's a Discord link in the badges if you want to ask questions.

Final Thoughts

Kody is solving a problem that's about to get a lot more common: as people run agents across multiple MCP hosts, the lack of shared state becomes painful fast. This project takes a sensible swing at it with a stack that's easy to deploy and a design that respects multi-tenancy from the start. It's clearly early—the README leans on internal docs for anything beyond the basics—but if you're building agents that need to remember things across tools, it's worth a look at the architecture docs before you roll your own.