← Public packages
@kentcdodds/kody-connect-talk
Kent C. Dodds's Cloudflare Connect 2026 talk: why your agents' stuff needs a common house, and how Kody builds one on Cloudflare.
Files
// Generated by `npm run build` from dist/index.html. Do not edit.
export const page = "<!doctype html>\n<html lang=\"en\">\n\t<head>\n\t\t<meta charset=\"UTF-8\" />\n\t\t<meta name=\"viewport\" content=\"width=device-width, initial-scale=1.0\" />\n\t\t<title>I Built Your Personal Software Ecosystem</title>\n\t\t<link rel=\"icon\" href=\"/__KODY_BASE__/deck/assets/img/kody-mark.png\" />\n\t\t<script>\n\t\t\ttry {\n\t\t\t\t// a new waiting-room order every load (the deck reads this seed)\n\t\t\t\tlocalStorage.setItem('kody-deck-seed', String(crypto.getRandomValues(new Uint32Array(1))[0]))\n\t\t\t} catch {}\n\t\t\ttry {\n\t\t\t\tconst t = localStorage.getItem('kody-deck-theme')\n\t\t\t\tif (t === 'light' || t === 'dark') document.documentElement.dataset.theme = t\n\t\t\t} catch {}\n\t\t\tif (new URLSearchParams(location.search).get('mode') === 'audience') {\n\t\t\t\tdocument.documentElement.classList.add('audience')\n\t\t\t\tdocument.title = 'Kody talk — Audience'\n\t\t\t}\n\t\t</script>\n\t\t<script type=\"module\" crossorigin src=\"/__KODY_BASE__/assets/index-8KgqVNyh.js\"></script>\n\t\t<link rel=\"stylesheet\" crossorigin href=\"/__KODY_BASE__/assets/index-670FTcGG.css\">\n\t</head>\n\t<body>\n\t\t<hyperframes-slideshow tabindex=\"0\">\n\t\t\t<hyperframes-player interactive src=\"/__KODY_BASE__/deck/index.html\"></hyperframes-player>\n\t\t\t<script type=\"application/hyperframes-slideshow+json\">\n{\n \"slides\": [\n {\n \"sceneId\": \"title\",\n \"notes\": \"While the room fills, the word Personal glitches into Team and back, and the subtitle types one terrible automation after another, in a new random order every time. (When it's time) Hi, I'm Kent. I teach people how to build software. Lately I've been building Kody. This talk is about giving every person, and every team, their own software ecosystem, and why Cloudflare is the only place I could build it.\",\n \"fragments\": [\n 1.33\n ]\n },\n {\n \"sceneId\": \"one-house\",\n \"notes\": \"Start with you and one agent. It's great. (Click) Over time it fills up with your stuff: memories, packages, triggers, connected accounts, apps. You're living in your agent's house.\",\n \"fragments\": [\n 3.4,\n 4.85\n ]\n },\n {\n \"sceneId\": \"switch\",\n \"notes\": \"Problem one. A better agent comes out, or you just want to try a different one. (Click) You can move. Your stuff can't. It's locked in the old house.\",\n \"fragments\": [\n 6.6,\n 8.45\n ]\n },\n {\n \"sceneId\": \"teammate\",\n \"notes\": \"Problem two. Your teammate lives in a different house, built by a different agent, full of their own stuff. (Click) You can't share what you built. Neither of you can come over.\",\n \"fragments\": [\n 10.225,\n 11.475\n ]\n },\n {\n \"sceneId\": \"common-house\",\n \"notes\": \"Kody is a common house. Everyone gets their own room for their own stuff. (Click) Any agent can walk in the front door: Claude Code, Codex, Grok Bot, whatever comes next. (Click) And you can share anything you've built with any teammate. (Click) And some things belong to the whole house: an org's packages, connections, and secrets, shared by everyone in it. Kody isn't another agent. It's the software platform underneath all of them.\",\n \"fragments\": [\n 13.075,\n 14.565,\n 16.215,\n 17.365\n ]\n },\n {\n \"sceneId\": \"roadmap\",\n \"notes\": \"Here's the rest of the talk. Every agent gets two tools: search and execute. (Click) They're available through three doors: MCP, the CLI, and a plain HTTP API. (Click) Then the things you keep in your room. For each one: the problem, how Kody solves it, and how it works under the hood on Cloudflare.\",\n \"fragments\": [\n 19.1,\n 20.3,\n 22.1\n ]\n },\n {\n \"sceneId\": \"search-card\",\n \"notes\": \"First door: search. Agents can't use what they can't find.\",\n \"fragments\": [\n 24.15\n ]\n },\n {\n \"sceneId\": \"search-problem\",\n \"notes\": \"MCP's default shape is: list every tool up front. Connect a few servers and your agent's context is full before you say hello. (Click) And more tools makes the model worse at picking the right one.\",\n \"fragments\": [\n 25.9,\n 27.15\n ]\n },\n {\n \"sceneId\": \"search-solution\",\n \"notes\": \"Kody exposes one search tool. The agent describes the task in plain words. (Click) It types what it wants... (Click) and gets back a short ranked list: built-in capabilities, your own packages, guides. Plus the exact call shape, so the next step can just run it.\",\n \"fragments\": [\n 28.825,\n 29.925,\n 31.215\n ]\n },\n {\n \"sceneId\": \"search-hood\",\n \"notes\": \"Under the hood: the query is embedded with Workers AI, matched in Vectorize filtered to your user id, and hydrated from D1. (Click) The same index holds your packages, guides, MCP servers, and secret names. Never secret values.\",\n \"fragments\": [\n 33.185,\n 34.905\n ]\n },\n {\n \"sceneId\": \"execute-card\",\n \"notes\": \"Second door: execute.\",\n \"fragments\": [\n 37.025\n ]\n },\n {\n \"sceneId\": \"execute-problem\",\n \"notes\": \"Classic tool calling is a round trip per step. The model calls a tool, the whole result comes back through the context, the model reads it and calls the next one. (Click) Four steps, four round trips, and every byte of every result goes through the model.\",\n \"fragments\": [\n 39.15,\n 40.75\n ]\n },\n {\n \"sceneId\": \"execute-solution\",\n \"notes\": \"With execute, the agent writes one small module and Kody runs it. Loops, filters, error handling: all in code, where models are already great. (Click) Every kody dot call is a capability it found with search. One round trip, and only the answer comes back.\",\n \"fragments\": [\n 42.325,\n 43.525\n ]\n },\n {\n \"sceneId\": \"execute-hood\",\n \"notes\": \"Every execute call gets a brand-new isolate from a Worker Loader. Milliseconds to start, nothing shared with the last run or anyone else. kody:runtime is a binding into Kody's capabilities. (Click) And the isolate has no network of its own. Every fetch goes through Kody's gateway, which checks the host and fills in secrets.\",\n \"fragments\": [\n 45.315,\n 46.785\n ]\n },\n {\n \"sceneId\": \"mcp-card\",\n \"notes\": \"Search and execute are the whole API surface. Three doors expose them: MCP, the CLI, and HTTP. MCP first, the one most of you already use.\",\n \"fragments\": [\n 48.9\n ]\n },\n {\n \"sceneId\": \"mcp-how\",\n \"notes\": \"MCP is one URL with OAuth: kody.codes/mcp. One command finds the agents on your machine and wires each one up. Claude Code, Codex, Cursor, whatever you use, they all get search and execute. Or skip that: some agents have Kody in their plugin marketplace, so it's one click. (Click) And my favorite trick: an agent that's already connected over MCP can mint a one-shot code for the CLI. The CLI redeems it for a scoped token in your keychain. No login flow, no pasting tokens.\",\n \"fragments\": [\n 52.175,\n 53.575\n ]\n },\n {\n \"sceneId\": \"cli-card\",\n \"notes\": \"The CLI is the door most people haven't seen. It runs execute somewhere surprising: your own machine.\",\n \"fragments\": [\n 55.4\n ]\n },\n {\n \"sceneId\": \"cli-problem\",\n \"notes\": \"Execute runs in Kody's cloud sandbox, which is exactly what you want most of the time. (Click) But the sandbox can't see your laptop: local files, a dev server on localhost, the NAS on your home network.\",\n \"fragments\": [\n 57.225,\n 57.975\n ]\n },\n {\n \"sceneId\": \"cli-solution\",\n \"notes\": \"So the CLI can run execute locally. Same module, same kody calls, same imports from your packages. (Click) It reads the CSV on your disk and still sends the email through Kody. Your CPU is free; only the Kody calls count.\",\n \"fragments\": [\n 59.775,\n 60.725\n ]\n },\n {\n \"sceneId\": \"cli-hood\",\n \"notes\": \"Your module runs in workerd, the same open-source runtime as Workers, on your machine. Every kody:runtime call becomes a request to the CapabilityProxy at api.kody.codes, which dispatches through the exact same tool map as cloud execute. (Click) Imports from your packages are downloaded as a package graph and bundled into the local worker. And a fetch that carries a secret placeholder hops to Kody's fetch gateway, so secret values never enter your machine.\",\n \"fragments\": [\n 62.685,\n 63.785\n ]\n },\n {\n \"sceneId\": \"api-card\",\n \"notes\": \"And the third door: plain HTTP.\",\n \"fragments\": [\n 65.9\n ]\n },\n {\n \"sceneId\": \"api-how\",\n \"notes\": \"The third door is a plain HTTP API, for anything that isn't an agent: scripts, CI, other apps. api.kody.codes is its own small Worker at the edge: CORS, rate limits, a body cap. It hands off to Kody over a service binding, no public hop, with a scoped token. (Click) Same search, same execute, same isolation, same audit trail, whatever the client.\",\n \"fragments\": [\n 67.935,\n 69.575\n ]\n },\n {\n \"sceneId\": \"memory-card\",\n \"notes\": \"Now what's in your room. Memory first.\",\n \"fragments\": [\n 71.65\n ]\n },\n {\n \"sceneId\": \"memory-problem\",\n \"notes\": \"Every agent keeps its own memory, and none of them can see the others'. So every new agent asks the same questions. What time zone are you in? (Click) And you answer them. Again.\",\n \"fragments\": [\n 73.625,\n 75.275\n ]\n },\n {\n \"sceneId\": \"memory-solution\",\n \"notes\": \"In Kody, memory belongs to you, not to the agent. Tell one agent you're on Mountain Time, once. (Click) Every agent connected as you gets it, surfaced automatically when it's relevant.\",\n \"fragments\": [\n 76.825,\n 78.415\n ]\n },\n {\n \"sceneId\": \"memory-hood\",\n \"notes\": \"Agents send a short memoryContext with normal search and execute calls: what they're working on. Kody matches it against your memories in Vectorize and inlines the best few. The memory itself lives in D1, scoped to you. (Click) Writes are verify-first: search for a related memory before saving a new one, so memory stays small and true.\",\n \"fragments\": [\n 80.435,\n 81.785\n ]\n },\n {\n \"sceneId\": \"packages-card\",\n \"notes\": \"Packages.\",\n \"fragments\": [\n 83.9\n ]\n },\n {\n \"sceneId\": \"packages-problem\",\n \"notes\": \"Your agent writes genuinely useful code all day long. (Click) Then the chat ends, and the code is gone. Next week you ask for the same thing and get a slightly different version.\",\n \"fragments\": [\n 85.775,\n 87.125\n ]\n },\n {\n \"sceneId\": \"packages-solution\",\n \"notes\": \"In Kody, that code becomes a package. A real Git repo with history, owned by you. Your own GitHub. (Click) And any code can import it by name, from execute, from another package, from an app. Your own npm.\",\n \"fragments\": [\n 88.7,\n 89.95\n ]\n },\n {\n \"sceneId\": \"packages-hood\",\n \"notes\": \"Every package is a Git repo on Cloudflare Artifacts. Agents edit in a repo session, a Durable Object holding a working copy. Publish runs checks: types, exports, imports. (Click) What ships is a bundle snapshot in KV, which a Worker Loader runs on demand. Static imports between packages are resolved at publish, so calling another package costs nothing extra.\",\n \"fragments\": [\n 91.815,\n 93.315\n ]\n },\n {\n \"sceneId\": \"triggers-card\",\n \"notes\": \"Triggers.\",\n \"fragments\": [\n 95.4\n ]\n },\n {\n \"sceneId\": \"triggers-problem\",\n \"notes\": \"Agents can wake up on a schedule now, and that's great. But every run hands the task back to a model, which reinterprets it from scratch. (Click) Same cron, three Mondays, three different answers. You pay for the model every run, and the schedule lives in that one agent's house.\",\n \"fragments\": [\n 97.075,\n 98.785\n ]\n },\n {\n \"sceneId\": \"triggers-solution\",\n \"notes\": \"In Kody, the agent writes the code once, and a trigger runs that same code every time. A cron schedule. A webhook URL. An email address that's really yours. Events from inside Kody. (Click) Each one runs your package, as you, with no model in the loop unless you put one there.\",\n \"fragments\": [\n 100.75,\n 102.1\n ]\n },\n {\n \"sceneId\": \"triggers-hood\",\n \"notes\": \"Schedules live in a JobManager Durable Object, using alarms to wake up exactly when the next job is due. Long jobs hand off to Workflows: durable steps that retry and survive restarts. (Click) Mail comes in through Email Routing into a Mailbox Durable Object per inbox, with the raw message in R2. Events fan out to your package subscriptions.\",\n \"fragments\": [\n 103.94,\n 105.41\n ]\n },\n {\n \"sceneId\": \"credentials-card\",\n \"notes\": \"Credentials.\",\n \"fragments\": [\n 107.525\n ]\n },\n {\n \"sceneId\": \"credentials-problem\",\n \"notes\": \"How do most of us give an agent access to an API today? We paste the key into the chat. (Click) Now it lives in the transcript, the model's context, maybe logs, forever. (Click) And if you're thinking 'I'd never do that'... your .env file is even worse. Every agent on your machine can read it, and it has all your keys, not just one.\",\n \"fragments\": [\n 109.2,\n 110.8,\n 112.05\n ]\n },\n {\n \"sceneId\": \"credentials-solution\",\n \"notes\": \"In Kody, you save the secret once, through a page the agent never sees. Code refers to it by name. (Click) The value is filled in on the way out the door, only for hosts you approved. The agent uses your keys without ever holding them.\",\n \"fragments\": [\n 113.7,\n 115\n ]\n },\n {\n \"sceneId\": \"credentials-hood\",\n \"notes\": \"The isolate only ever sees the placeholder. Its outbound fetch is routed to Kody's fetch gateway, which checks the host against the secret's allowlist, decrypts the value, and swaps it in. (Click) Send the same placeholder anywhere else, and the request is refused before it leaves.\",\n \"fragments\": [\n 116.935,\n 118.285\n ]\n },\n {\n \"sceneId\": \"apps-card\",\n \"notes\": \"Apps.\",\n \"fragments\": [\n 120.4\n ]\n },\n {\n \"sceneId\": \"apps-problem\",\n \"notes\": \"Chat is great for asking. It's a terrible control center. You can read the state, sort of, as a wall of JSON. (Click) And acting on it means typing commands back and hoping you picked the right one. Some work wants a UI you can come back to, see at a glance, and click.\",\n \"fragments\": [\n 122.075,\n 123.325\n ]\n },\n {\n \"sceneId\": \"apps-solution\",\n \"notes\": \"So packages can ship apps. Not just a dashboard: a control center. See every PR at a glance. (Click) And act: merge the green one, turn on auto-merge. Every button calls your package's code, as you. (Click) It lives at your own subdomain on kody.run, or on your own domain, and you can share it with your team like anything else in Kody.\",\n \"fragments\": [\n 124.95,\n 126.6,\n 127.4\n ]\n },\n {\n \"sceneId\": \"apps-hood\",\n \"notes\": \"A request to your subdomain loads your app's fetch handler into a Worker Loader isolate, the same sandbox as execute. Its data lives in a StorageRunner Durable Object: a private SQLite database per storage bucket. (Click) Need realtime? Each session is a Durable Object holding the WebSockets.\",\n \"fragments\": [\n 129.19,\n 130.54\n ]\n },\n {\n \"sceneId\": \"sharing-card\",\n \"notes\": \"Back to the house. Sharing.\",\n \"fragments\": [\n 132.65\n ]\n },\n {\n \"sceneId\": \"sharing-solution\",\n \"notes\": \"Remember the teammate in the other house? In Kody, sharing is one move: grant a person or a team a role on a resource. A package, an app, a secret, a connection. (Click) And because it's all in one house, whatever agent your teammate uses can use it too.\",\n \"fragments\": [\n 134.735,\n 136.335\n ]\n },\n {\n \"sceneId\": \"sharing-hood\",\n \"notes\": \"How do you keep everyone's room private in one shared house? Every layer is namespaced by owner. D1 rows carry the owner id. Durable Objects are named per owner. Vectorize queries filter by owner. KV and R2 keys are prefixed. (Click) And every request, from every surface, goes through one authorize check. One door policy for the whole house.\",\n \"fragments\": [\n 138.3,\n 139.6\n ]\n },\n {\n \"sceneId\": \"cloudflare\",\n \"notes\": \"Per-person isolates on demand, per-person databases by the million, per-person Git, email, queues, and vector search. All in one platform, all billed by actual use. That's why Kody could only be built on Cloudflare. (Click) And all of it is available to you today.\",\n \"fragments\": [\n 142,\n 143.35\n ]\n },\n {\n \"sceneId\": \"invisible-software\",\n \"notes\": \"This is the scale I'm building for. Today, tens of millions of developers are creating code with agents. (Click) I think we're heading toward billions of people whose agents create software for them. They don't need to become developers, or even realize software was written. They just want their problems solved. (Click) I call this Invisible Software. That's why I built Kody with this scale in mind. My latest video goes deeper into that vision; the thumbnail and link open it. Then: thank you, and come build with us.\",\n \"fragments\": [\n 145.25,\n 146.2,\n 147.35\n ]\n },\n {\n \"sceneId\": \"thanks\",\n \"notes\": \"Thank you! Scan the code or go to kody.codes. Connect it to the agent you already use. I would love to hear what you build.\",\n \"fragments\": [\n 149.6\n ]\n }\n ]\n}\n</script>\n\t\t</hyperframes-slideshow>\n\n\t\t<div class=\"theme-toggle\" role=\"radiogroup\" aria-label=\"Color theme\">\n\t\t\t<button type=\"button\" role=\"radio\" data-theme-choice=\"light\" aria-label=\"Light theme\" title=\"Light\">\n\t\t\t\t<svg viewBox=\"0 0 24 24\"><circle cx=\"12\" cy=\"12\" r=\"4\" /><path d=\"M12 2v2M12 20v2M4.9 4.9l1.4 1.4M17.7 17.7l1.4 1.4M2 12h2M20 12h2M4.9 19.1l1.4-1.4M17.7 6.3l1.4-1.4\" /></svg>\n\t\t\t</button>\n\t\t\t<button type=\"button\" role=\"radio\" data-theme-choice=\"system\" aria-label=\"Match system theme\" title=\"System\">\n\t\t\t\t<svg viewBox=\"0 0 24 24\"><rect x=\"3\" y=\"4\" width=\"18\" height=\"12\" rx=\"2\" /><path d=\"M8 20h8M12 16v4\" /></svg>\n\t\t\t</button>\n\t\t\t<button type=\"button\" role=\"radio\" data-theme-choice=\"dark\" aria-label=\"Dark theme\" title=\"Dark\">\n\t\t\t\t<svg viewBox=\"0 0 24 24\"><path d=\"M20.5 14.5A8.5 8.5 0 0 1 9.5 3.5a8.5 8.5 0 1 0 11 11z\" /></svg>\n\t\t\t</button>\n\t\t</div>\n\n\t</body>\n</html>\n"