Cloudflare
The Cloudflare target builds your agents for the Cloudflare platform. Each agent runs inside its own Durable Object class, using the Agents SDK, Workers AI, Cloudflare Sandbox, Cloudflare Shell, and other Worker primitives where appropriate. Durable Objects give each agent conversation its own persistent state, durable execution, and global addressability out of the box.
The build is owned by Vite: flue() plus the official @cloudflare/vite-plugin in vite.config.ts, with flue() first. See Deploy for the plugin mechanics and Deploy Agents on Cloudflare for a deployment walkthrough.
Generated Durable Objects
Flue generates a Durable Object class and a Wrangler binding for each agent found by the 'use agent' scan — one per exported agent function, so a module with several agents produces several classes. Both derive from the agent’s identity — the exported function’s name, or its agentName static override:
export function SupportChat() {…} -> FlueSupportChatAgent
env.FLUE_SUPPORT_CHAT_AGENTCamel boundaries split in the binding name, so SupportChat yields FLUE_SUPPORT_CHAT_AGENT (never FLUE_SUPPORTCHAT_AGENT); a kebab-case agentName like support-chat generates the same pair. The class name is how Cloudflare identifies the Durable Object in migrations. The binding is how your application code accesses the Durable Object namespace at runtime through env. Because both come from the identity, renaming the agent function is a storage-identity change unless agentName pins the identity — renaming the file changes nothing; see Managing migrations.
Canonical agent conversation streams, immutable attachments, and accepted submissions are stored in the owning Durable Object’s SQLite storage automatically. The Cloudflare target does not use db.ts; a source-root db.ts is rejected at build time.
Do not hand-author Flue’s generated FLUE_* bindings in wrangler.jsonc. Declare migrations for generated classes, and declare bindings only for application-owned resources such as your own Durable Objects, R2 buckets, Queues, Hyperdrive configs, Browser Rendering bindings, or Send Email bindings.
wrangler.jsonc
Your project’s wrangler.jsonc at the project root configures your Worker’s name, compatibility settings, and Durable Object migrations. Flue reads this file, merges its contributions (main and the per-agent Durable Object bindings) into a generated .flue-vite.wrangler.jsonc that the Cloudflare Vite plugin consumes, and never modifies your authored file. Add .flue-vite/ and .flue-vite.wrangler.jsonc to .gitignore.
Flue generates the Durable Object classes and bindings, but your wrangler.jsonc must declare two things:
nodejs_compatincompatibility_flags, because Flue’s runtime uses Node.js APIs.- Durable Object migrations that list every generated class. Cloudflare requires an explicit migration whenever a Worker adds, renames, or removes a Durable Object class.
{
"$schema": "./node_modules/wrangler/config-schema.json",
"name": "my-flue-worker",
"compatibility_date": "2026-06-01",
"compatibility_flags": ["nodejs_compat"],
"migrations": [
{
"tag": "v1",
"new_sqlite_classes": ["FlueSupportChatAgent"],
},
],
}Managing migrations
Migration history stays user-authored — Flue never writes it, because it is an ordered, append-only record of your deployments. Adding an agent is always a triple: the exported agent function in a 'use agent' module, the mount in app.ts (unless dispatch-only), and a new migration entry with a unique tag:
{
"migrations": [
{ "tag": "v1", "new_sqlite_classes": ["FlueSupportChatAgent"] },
{ "tag": "v2", "new_sqlite_classes": ["FlueTriageAgent"] },
],
}Never rewrite or reorder deployed migration entries. Generated agent classes require Durable Object SQLite, so introduce them through new_sqlite_classes, not legacy new_classes. Use Cloudflare’s renamed_classes and deleted_classes migration fields when changing deployed class names or removing classes.
For example, if you remove an agent that was previously deployed, append a deleted_classes migration so Cloudflare knows the class is no longer exported. Without this entry, Wrangler will fail because the migration history references a class that the Worker no longer provides:
{
"migrations": [
{ "tag": "v1", "new_sqlite_classes": ["FlueSupportChatAgent", "FlueTriageAgent"] },
{ "tag": "v2", "deleted_classes": ["FlueTriageAgent"] },
],
}Similarly, use renamed_classes when a deployed class changes its name — which for Flue agents means the agent’s identity changed: the function was renamed, or its agentName pin was edited:
{
"migrations": [
{ "tag": "v1", "new_sqlite_classes": ["FlueSupportChatAgent"] },
{
"tag": "v2",
"renamed_classes": [{ "from": "FlueSupportChatAgent", "to": "FlueSupportAssistantAgent" }],
},
],
}Renaming preserves the class’s stored conversations under the new name. Re-mounting an agent at a different URL in app.ts, by contrast, needs no migration at all — the mount path is not part of the storage identity.
Local development
vite dev runs your Worker in workerd via the Cloudflare plugin, with the same permissive dev CORS defaults as the Node target (Vite’s middleware stack applies them before workerd sees the request). Flue keeps the generated inputs fresh: agent-set changes (adding or removing a 'use agent' file, or an agent export within one) and authored wrangler.jsonc edits regenerate the entry and merged config, and the Cloudflare plugin picks them up. Writes are content-aware — editing an agent’s body regenerates nothing, so there are no restart loops. vite preview and deployment come from the Cloudflare plugin; see Deploy Agents on Cloudflare.
The plugin registers agents but mounts no routes. Your app.ts remains the route map.
Durable agent execution
Cloudflare agents durably admit direct HTTP prompts together with dispatch(...) inputs. All accepted input for one agent conversation enters the same queue.
direct HTTP prompt ─────────────────────┐
├→ durable per-conversation queue → canonical stream
dispatch(...) input ────────────────────┘The submitting connection observes the work but does not own it. If a client disconnects after admission, backend work can continue. Agent events are durably stored and can be replayed from any offset via the Durable Streams protocol.
When a Durable Object resumes after interruption, Flue decides what to do next from the stored input and canonical conversation progress. It requeues only when it can prove the input was not applied, recognizes already-completed output, and records an interruption instead of blindly repeating uncertain model or tool work.
For the full recovery model, see Durability.
Calling a private agent over a service binding
A Flue Worker deployed without a public route can still be reached from another Worker through a service binding. The SDK client sends every request through its fetch option, so point that option at the binding instead of the network:
import { createFlueClient } from '@flue/sdk';
type Env = { AGENT_APP: Fetcher };
export default {
async fetch(request: Request, env: Env) {
const convo = createFlueClient({
// The host is never dialed — only the pathname and query select a route.
// The URL must be absolute, so any placeholder origin works; the path is
// wherever the agent app's app.ts mounts the agent, plus the conversation id.
url: 'https://agent.internal/agents/support/ticket-42',
fetch: (input, init) => env.AGENT_APP.fetch(new Request(input, init)),
});
const admission = await convo.send({
message: { kind: 'user', body: 'Summarize this ticket.' },
});
return Response.json(admission);
},
};The binding carries the same HTTP requests the public routes use, so every client operation works over it — send, wait, abort, history, and observe in both long-poll and sse modes. Streaming reads travel through the same fetch, so live conversation updates cross the binding and the owning Durable Object with no extra wiring.
Attachments are the one exception. convo.attachmentUrl(...) returns a URL on the placeholder host and the client never fetches it for you; the same URL also appears on file parts in observe() and history() snapshots. To download attachment bytes over a binding, forward that URL through the same fetcher:
const url = convo.attachmentUrl(attachmentId);
const response = await env.AGENT_APP.fetch(new Request(url));Workers AI and AI Gateway
Workers AI lets you run AI models directly on Cloudflare’s infrastructure without managing API keys or external provider accounts. Flue connects to Workers AI automatically on the Cloudflare target, so using a Workers AI model is as simple as specifying the model name:
export function Assistant() {
useModel('cloudflare/@cf/moonshotai/kimi-k2.6');
}No API key is needed. Authorization and billing follow the Worker account, including the Workers AI free tier.
Flue also enables AI Gateway by default for all cloudflare/... models, giving you caching, request logging, rate limiting, and budget controls in the Cloudflare dashboard out of the box.
To customize the gateway, disable it, or target a named gateway, re-register the cloudflare provider in app.ts. See Cloudflare Workers AI for examples.
Cloudflare Sandbox
Cloudflare Sandbox provides container-backed Linux environments for agents that need tools such as git, package installation, native binaries, or a real filesystem. Export the sandbox Durable Object class from cloudflare.ts, declare its binding and container image in wrangler.jsonc, then wrap the RPC stub returned by getSandbox(...) with cloudflareSandbox(...):
'use agent';
import { getSandbox } from '@cloudflare/sandbox';
import { env } from 'cloudflare:workers';
import { type AgentProps, useModel, useSandbox } from '@flue/runtime';
import { cloudflareSandbox } from '@flue/runtime/cloudflare';
export function Assistant({ id }: AgentProps) {
useModel('anthropic/claude-sonnet-4-6');
useSandbox(cloudflareSandbox(getSandbox(env.Sandbox, id)), { cwd: '/workspace' });
}See Cloudflare Sandbox for container configuration and lifecycle guidance.
Codemode
By default, Flue agents use a lightweight in-memory virtual sandbox. This is fast and sufficient for prompt-and-response agents or agents that only need tools and structured results. When an agent needs a durable workspace with structured code execution instead of a full Linux container, use Cloudflare Shell with Codemode.
Cloudflare Shell provides a durable Workspace with a model-facing code tool backed by @cloudflare/codemode. The agent interacts with files through structured code operations rather than shell commands. This means harness.sandbox.exec(...) does not run arbitrary Linux commands through this sandbox adapter — it throws, since the Workspace has no shell.
Add the sandbox adapter to your project:
pnpm exec flue add sandbox cloudflare-shellThen import its helpers from your generated sandbox adapter file, not from @flue/runtime/cloudflare:
import { getDefaultWorkspace, getShellSandbox } from '../sandboxes/cloudflare-shell';Use Cloudflare Shell when a durable Workspace and structured code operations are enough. Use Cloudflare Sandbox when you need a full Linux environment with arbitrary shell access. See Cloudflare Shell for setup details.
Extending Agents on Cloudflare
Flue owns each generated Durable Object class. When an agent needs access to native Cloudflare Agents SDK capabilities such as onStart(), schedule(), scheduleEvery(), or queue(), export a cloudflare extension descriptor from its module:
'use agent';
import { useModel } from '@flue/runtime';
import { extend } from '@flue/runtime/cloudflare';
export function Assistant() {
useModel('anthropic/claude-sonnet-4-6');
}
export const cloudflare = extend({
base: (Base) =>
class extends Base {
async onStart() {
await this.scheduleEvery(60, 'heartbeat');
}
async heartbeat() {
this.setState({ ...this.state, lastHeartbeatAt: Date.now() });
}
},
});base receives the Agents SDK Agent base class. Flue applies it before defining the final generated Durable Object subclass, so your authored methods and lifecycle hooks are available on the generated class. The cloudflare export is per-module: when a marked file exports several agents, the extension applies to every generated class in that file.
wrap receives the final generated class and may return a prototype-preserving constructor wrapper. Use it for integrations like Sentry that instrument the class without replacing its prototype:
export const cloudflare = extend({
wrap: (Final) =>
Sentry.instrumentDurableObjectWithSentry((env) => ({ dsn: env.SENTRY_DSN }), Final),
});Both base and wrap are optional. Do not override Flue-owned fetch(), onRequest(), onFiberRecovered(), or alarm() methods.
Use this module-local extension point for scheduled or queued behavior that belongs to one generated agent Durable Object. Do not add a Worker cron trigger just to reach scheduleEvery(...); the Agents SDK scheduling APIs run inside the generated Durable Object after that object is created. If your application needs to create the first conversation, expose an authenticated bootstrap route in app.ts or otherwise obtain the Durable Object namespace from env and address the conversation once.
Extending cloudflare.ts Entrypoint
Your project may include a source-root cloudflare.ts file for Worker-level Cloudflare code that is separate from individual agent modules.
Any named export from this file becomes a top-level Worker export. This is how you add application-owned Durable Objects to the same Worker that Flue manages. For example, a cache Durable Object that your agents can access through env:
import { DurableObject } from 'cloudflare:workers';
// This class becomes a Worker export. Declare its binding and
// migration in wrangler.jsonc so Cloudflare knows about it.
export class SalesforceAuthCache extends DurableObject {
async refreshIfNeeded() {
return await this.ctx.storage.get('token');
}
}After exporting the class, declare its Durable Object binding and migration in wrangler.jsonc. Your agents can then access it through env.SALESFORCE_AUTH_CACHE.
The default export may contribute non-HTTP Worker handlers. For example, a scheduled handler that runs on a cron trigger:
export default {
async scheduled(_controller, env) {
await env.SALESFORCE_AUTH_CACHE.getByName('default').refreshIfNeeded();
},
};Use app.ts for custom HTTP routes and middleware. cloudflare.ts must not define a default fetch handler because Flue keeps HTTP composition in app.ts.
Use cloudflare.ts for Worker-level events such as inbound email, queues, or cron handlers that are not owned by a specific generated agent class. To deliver scheduled input to an agent from one of these handlers, import the agent function and call dispatch(Agent, { id, message }) — dispatch needs no mount and bypasses HTTP middleware. See Schedules for a Cron Trigger example.
Reference
extend(...)
import { extend } from '@flue/runtime/cloudflare';
function extend<TBase extends object = CloudflareAgentLike, TEnv = any>(
extension: CloudflareExtension<TBase, TEnv>,
): CloudflareExtension<TBase, TEnv>;Creates a branded Cloudflare extension descriptor for an agent module. The descriptor may contain base and wrap callbacks.
Both callbacks are typed against CloudflareAgentLike, a structural view of the Agents SDK Agent base class covering state, setState(), onStart(), schedule(), scheduleEvery(), and queue(), so typos inside base callbacks fail at typecheck. Pass an explicit TBase (for example extend<CloudflareAgentLike<MyState>>({ ... })) to type against a richer class shape, and an explicit TEnv to type the env an instrumentation callback receives.
base(Base) must return the received class or a subclass. Flue uses its return value as the superclass for the generated Durable Object.
wrap(Final) must return the received class or a prototype-preserving constructor wrapper. Use it for integrations that instrument or proxy the final generated class without replacing its prototype. Subclasses are rejected; only the same class or a new Proxy(Final, {...}) pattern is allowed. The class both callbacks receive is typed as a real Durable Object constructor, so brand-checked wrappers such as @sentry/cloudflare’s instrumentDurableObjectWithSentry accept it directly, with no casts or explicit generics.
Both callbacks are optional. When omitted, the corresponding step is an identity operation.
getCloudflareContext()
import { getCloudflareContext } from '@flue/runtime/cloudflare';
function getCloudflareContext(): CloudflareContext;Returns the current Cloudflare runtime context. Only valid while code is running inside a Worker or Durable Object request handler.
The returned CloudflareContext includes:
env– the Worker’s environment bindings.storage– the Durable Object’s{ sql }SQLite storage handle.
Throws outside of Cloudflare runtime work.
This is intended for advanced application-owned integrations such as custom Cloudflare sandbox adapters. Most applications do not need to call this directly.
getDurableObjectIdentity()
import { getDurableObjectIdentity } from '@flue/runtime/cloudflare';
function getDurableObjectIdentity(): FlueDurableObjectIdentity;Returns the generated Durable Object identity for the current agent context. Only valid inside a generated Durable Object request handler.
The returned FlueDurableObjectIdentity includes:
bindingName– the Wrangler binding name, such as"FLUE_SUPPORT_CHAT_AGENT".className– the generated class name, such as"FlueSupportChatAgent".name– the instance name passed toidFromNameorgetAgentByName.id– the Durable Object ID as a string.
Throws when called outside a generated Durable Object context.