kody-hq/kody

🐨 The cloud for your team's agents. Save your agent memory, credentials, code, and automations, portable across every agent with Kody.

Kody – Cloud‑native “cloud” for AI agents

What it is – Kody is a server‑side platform that stores the state of AI agents: memory fragments, API keys, code packages, and scheduled jobs. It exposes this data through the Model Context Protocol (MCP) so that any MCP‑compatible host (e.g., ChatGPT plugins) can read or write an agent’s context. The whole system runs on Cloudflare Workers, making it globally distributed, low‑latency, and cheap to operate.

Why it matters – Modern LLM‑based agents need a place to keep persistent information (past conversations, tool definitions, credentials) that survives across invocations. Kody provides that persistence as a first‑class service, with built‑in OAuth protection and a Remix‑based UI for humans to inspect or edit the stored data.

Core components

Component	Role
kody‑worker (origin)	Entry point that serves the Remix UI, handles OAuth, and routes MCP HTTP requests.
kody‑platform‑worker	Implements the MCP mailbox, metering, and session state using Cloudflare Durable Objects.
kody‑runtime‑worker	Executes user‑uploaded packages (code, tools) on demand via the `execute` flow.
kody‑jobs‑worker	Cron‑style job scheduler; runs background tasks and reports back through the JobsHost.
mock‑servers	Local test doubles for Workers, enabling fast CI/E2E runs.

Key features

MCP‑first API – A compact set of endpoints (search, execute, etc.) that any MCP‑compatible AI host can call.
Persistent memory – Durable Objects store per‑account “memories” that agents can query or update.
Secure credentials – OAuth‑protected endpoints and Cloudflare KV store API keys/secret tokens.
Package runtime – Users can upload code packages that the runtime worker can invoke safely.
Job scheduling – Built‑in cron manager for recurring tasks (e.g., data refresh, cleanup).
Team isolation – Each account (individual or organization) gets its own isolated namespace for packages, jobs, credentials, and memories.
Local dev experience – wrangler spins up a full Workers + D1 (SQLite‑compatible) stack on localhost:3742.

Tech stack (as listed in the README)

Runtime: Cloudflare Workers (serverless edge platform)
UI: Remix 3 (beta) bundled with Vite
Package manager: npm (Nx monorepo workspace)
Database: Cloudflare D1 (SQL‑like edge DB)
Session / OAuth storage: Cloudflare KV
Stateful MCP objects: Cloudflare Durable Objects
Testing: Playwright for end‑to‑end tests

Typical workflow

Create an account – via the UI or API; OAuth secures the session.
Upload a package – code that defines tools or custom logic for the agent.
Add memories / credentials – stored in Durable Objects / KV.
Agent host (e.g., ChatGPT) calls MCP – search to retrieve relevant memories, execute to run a package.
Background jobs – optional cron jobs run in the jobs worker and can update memories or trigger external services.

Getting started (from the README)

# clone the repo, install deps
npm install
# start the local development server (Workers + D1 are auto‑started)
npm run dev

The dev server prints a URL like http://localhost:3742. For a fresh project based on the same conventions you can also run npx create‑epicflare.

Documentation highlights

Architecture – docs/contributing/architecture/index.md explains the four‑worker topology.
Project intent – docs/contributing/project-intent.md clarifies scope (MCP‑first, not a generic tool catalog).
Usage guide – docs/use/index.md shows how an external MCP host should interact with Kody.
Setup – docs/contributing/getting-started.md and setup folder cover env vars, Cloudflare bindings, and local verification.

License

Kody is released under the Functional Source License 1.1 (FSL‑1.1‑ALv2), which permits free use, modification, and redistribution except for “Competing Use” (commercial products that replicate Kody’s core functionality). Two years after a version’s release it re‑licenses to Apache 2.0.

Who might use it?

LLM‑powered agents that need a persistent, searchable memory store.
Developers building ChatGPT plugins or other MCP‑compatible services.
Teams that want a shared, secure place for agent credentials, code, and scheduled jobs.
Researchers experimenting with agent‑state management without building their own edge infrastructure.

Built with ❤️ by Epic Web.