Workers fetch() requests now support the cf.vary request option. Use cf.vary to control how Cloudflare caches origin responses with a Vary header for a single subrequest.
The latest release of the Agents SDK ↗ makes it easier to run long work in the background, drive turns through one entry point, and keep chat agents working through deploys, evictions, and reconnects.
This release adds first-class detached (background) sub-agent runs with live progress and durable milestones, a single runTurn turn-admission entry point, and a large round of recovery and reliability fixes that continue converging @cloudflare/think and @cloudflare/ai-chat onto one model.
Background sub-agents with progress and milestones
runAgentTool can now dispatch a sub-agent without blocking the calling turn. A detached run returns a handle immediately and is owned by a durable, eviction-surviving backbone instead of being abandoned when the dispatching turn ends.
Durable, exactly-once-on-the-happy-path completion via a warm fast path plus a self-scheduling reconcile backbone that survives eviction and deploys.
Bounded. An absolute maxBudgetMs ceiling (default 24h) and cancelAgentTool(runId) keep abandoned runs from holding a concurrency slot forever.
detached: { notify: true } lets a finished background run inject a message back into the chat so the model reacts to the result — no hand-wired onFinish needed.
Sub-agents can also report mid-run progress that rides their own turn stream back to the parent's connected clients:
// Inside the child sub-agent:await this.reportProgress({ fraction: 0.6, phase: "deploying", message: "Generating menu page…",});
// Inside the child sub-agent:await this.reportProgress({ fraction: 0.6, phase: "deploying", message: "Generating menu page…",});
Progress surfaces on AgentToolRunState.progress via useAgentToolEvents, so a background-runs tray can render a live bar without drilling in, and the latest snapshot is persisted for inspection after eviction. Naming a milestone promotes a signal to a durable, replayable row, and detached: { onMilestones } can surface a milestone as a synthetic chat message ("narrate" for a cheap status line, or "react" to drive a model turn).
One entry point for turns: runTurn
@cloudflare/think adds a public runTurn(options) facade that unifies turn admission behind a single mode:
stream mode accepts array and function inputs to match wait mode, and all entry points now route through a shared internal admission path that throws a clear error on nested blocking admissions that previously could deadlock.
Recovery and reliability
A large part of this release continues hardening recovery and converging @cloudflare/think and @cloudflare/ai-chat onto one model:
Stream stall watchdog.AIChatAgent can detect and recover from a hung model/transport stream via the opt-in chatStreamStallTimeoutMs watchdog. With chatRecovery enabled the stall routes into the same bounded-recovery machinery a deploy or eviction uses; otherwise it surfaces as a terminal stream error so the spinner clears.
Interrupted tool-call repair.AIChatAgent now repairs a transcript with a dead server-tool call before re-entering inference (parity with @cloudflare/think), so a recovered turn no longer fails with AI_MissingToolResultsError. An overridable repairInterruptedToolPart(part) hook lets apps customize the repaired shape.
Stuck status after reconnect. Fixed AI SDK status getting stuck when a reconnect races a turn that has been accepted but has not started streaming yet, so the UI now renders the in-flight turn instead of settling on ready.
Live "recovering…" on connect.AIChatAgent now replays the recovering status to a client that connects mid-recovery, so useAgentChat's isRecovering reflects in-progress recovery immediately instead of appearing frozen.
Terminal connection failures. The client stops reconnecting on terminal WebSocket close events and exposes them via connectionError / onConnectionError on AgentClient, useAgent, and useAgentChat.
Agent-tool child recovery. A healthy long-running sub-agent run is no longer abandoned as interrupted after a deploy (both @cloudflare/think and AIChatAgent).
Workflows from sub-agent facets. Agent Workflows can now start from sub-agent facets, with callbacks and Workflow RPC routed back to the originating facet.
Plus forward-progress crediting convergence, broadcast-first give-up ordering, an event-driven auto-continuation barrier, and structured row-size compaction in AIChatAgent.
Other improvements
Shared chat React core. A new agents/chat/react entry exposes useAgentChat, transport helpers, and shared wire types, with syncMessagesToServer for server-authoritative transcript storage. @cloudflare/think/react and @cloudflare/ai-chat/react are now thin wrappers over it.
Optional ai peer. The root agents and @cloudflare/codemode runtimes no longer reference AI SDK types, so they bundle without ai / zod installed; AI-specific entry points still require the peer when imported. just-bash likewise moves to an optional peer used only by the skills bash runner.
Code Mode. The default DynamicWorkerExecutor timeout increases from 30s to 60s, executions now dispose the dynamically-loaded Worker and its RPC stub after each run (fixing a flaky isolate-shutdown assertion), connector imports are cleaned up, and the outer MCP tool-call context is passed to openApiMcpServer request callbacks.
Voice. Voice turns now support AI SDK fullStream responses (and warn when textStream is used).
MCP.McpAgent server-to-client requests can now be sent from callbacks that do not inherit the agent's async context, including callbacks reached through Worker Loader RPC.
Experimental: server actions and channels. This release lays groundwork for guarded server actions (action() / getActions() with a durable replay ledger and approvals) and a unified channels surface (configureChannels(), deliverNotice()). Both are experimental and their APIs may change, so we don't recommend depending on them yet.
Upgrade
To update to the latest version:
npm i agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest @cloudflare/codemode@latest @cloudflare/voice@latest
You can now connect autonomous agents and bots to an MCP server portal using an Access service token. Service token sessions can reach upstream MCP servers through the portal without a browser-based OAuth flow.
To set this up:
Add a Service Auth policy that matches your service token to the portal's Access application.
Add a Service Auth policy that matches the same token to each linked MCP server's Access application.
Turn Require user auth off (on_behalf: false) for each linked server so the portal uses the admin credential instead of a per-user OAuth grant.
The bot connects with CF-Access-Client-Id and CF-Access-Client-Secret headers and sees the tools from every linked server it is authorized for. Servers that still require per-user OAuth are excluded from service token sessions because a service token cannot complete a per-user OAuth grant.
Durable Objects now supports a usjurisdiction, letting you create Durable Objects that only run and store data within the United States. Use the us jurisdiction when you need to keep a Durable Object's compute and storage inside the United States to meet data residency requirements.
Create a namespace restricted to the us jurisdiction the same way as any other jurisdiction:
Workers may still access Durable Objects constrained to the us jurisdiction from anywhere in the world. The jurisdiction constraint only controls where the Durable Object itself runs and persists data.
The @cloudflare/vitest-pool-workers package now includes evictDurableObject and evictAllDurableObjects test helpers, exported from cloudflare:test.
These helpers let you test how a Durable Object behaves across evictions, simulating the production lifecycle where an idle Durable Object can be evicted from memory.
import { evictDurableObject, evictAllDurableObjects } from "cloudflare:test";import { env } from "cloudflare:workers";const id = env.COUNTER.idFromName("my-counter");const stub = env.COUNTER.get(id);// Evict the Durable Object instance pointed to by a specific stubawait evictDurableObject(stub);// Close WebSockets instead of hibernating themawait evictDurableObject(stub, { webSockets: "close" });// Evict all currently-running Durable Objects in evictable namespacesawait evictAllDurableObjects();
These helpers are available in @cloudflare/vitest-pool-workers@0.16.20 and later.
This beta release introduces upgraded security of device registration to be hardware-backed. Registration tokens can now be generated in the Secure Enclave whenever available to provide stronger protection against device impersonation.
Additional changes and improvements
This release also introduces multiple fixes and improvements including:
Improved accessibility by using high contrast colors and more defined color boundaries when high contrast is enabled in the macOS Display settings.
Path MTU Discovery (PMTUD) is now enabled by default.
Fixed an issue where DNS queries would fail after the connection was idle, requiring users to retry.
Users can now register with team names in any case format without errors.
New UI fixes
Fixed an issue where users with invalid MDM configurations were returned to the onboarding screen after successful authentication.
Added a re-auth button and banner to the home screen so users don't miss it when their session expires.
Added clear error messaging when the Cloudflare certificate needs to be installed.
Brought back support for pausing the tunnel when connected to user-specified Wi-Fi networks for consumer users.
New client UI now surfaces Split tunnel configuration and Local Domain Fallback configuration.
Added ability to configure proxy mode for consumer users.
Added back the option to quit for consumer users.
Known issues
Registration may hang at "Checking your organization configuration" due to IPC errors. A system reboot should resolve the error, allowing registration to proceed.
AI Search now gives you more control over similarity cache freshness. Similarity cache helps reduce latency and inference cost by reusing responses for semantically similar queries.
With these updates, you can choose how long responses are eligible for reuse and clear cached responses when they may be stale.
Cache duration now defaults to 48 hours
Previously, AI Search cached responses for a fixed duration of 30 days. Cached responses now use the instance's cache_ttl setting, and the default is 48 hours.
You can set cache_ttl when creating or updating an instance to choose a cache duration from 10 minutes to 6 days.
Use a shorter TTL when your source content changes frequently and freshness is more important. Use a longer TTL when your content is stable and you want more cache reuse.
For example, set cache_ttl to 518400 to retain cached responses for 6 days:
{ "cache_ttl": 518400}
Purge cached responses
You can also purge all cached responses for an instance on demand. Purging cached responses does not delete indexed content or source files.
It prevents AI Search from reusing previous cached responses, so subsequent similar queries generate fresh answers and repopulate the cache.
curl -X POST "https://api.cloudflare.com/client/v4/accounts/$ACCOUNT_ID/ai-search/instances/$INSTANCE_NAME/purge_cache" \ -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN"
You can also purge cached responses from the instance settings page in the Cloudflare dashboard.
Refer to similarity cache for the full list of supported cache_ttl values and more details about cache behavior.
You can now, as an Organization Super Administrator, view organization-level audit logs in the Cloudflare dashboard, in addition to the existing API access.
Organization audit logs help you monitor activity across your organization. You can see who performed an action, what changed, when it happened, how it was performed, and whether it succeeded or failed.
You can filter and search logs by actor, action, result, resource, request details, and timestamp. Use these logs to troubleshoot changes, investigate unexpected access, and support security or compliance workflows.
If you are viewing account-level audit logs and the account belongs to an organization where you are an Organization Super Administrator, select View Organization Audit Logs to open the parent organization's audit logs.
To get started, go to Organizations, select your organization, then go to Manage Organization > Audit Logs.
WebSocket Analytics: A new dataset with fields including BytesReceivedClient, BytesReceivedOrigin, BytesSentClient, BytesSentOrigin, ClientASN, ClientIP, ClientRequestHost, ClientRequestPath, ClientRequestUserAgent, ColoCode, ConnectionCloseReason, ConnectionCloseSource, ConnectionID, ConnectionTransportCloseCode, EdgeEndTimestamp, EdgeStartTimestamp, and RayID.
Updated fields in existing datasets
Firewall events (added): ZoneName. The Firewall events dataset is now also available for account-scope Logpush, in addition to the existing zone scope.
Email Security Alerts (added): BCC, DKIMResult, DMARCPolicy, DMARCResult, and SPFResult.
For the complete field definitions for each dataset, refer to Logpush datasets.
Radar now plots your IPv4 and IPv6 locations on the IP page ↗, shows the Cloudflare data centers serving your connection, and includes more detail about the autonomous system (AS) your primary IP belongs to.
Your IP location on the map
The map of your connection now shows:
IP location markers — The primary IP will show as a red marker. When both IP addresses do not geolocate to the same place, a second marker will appear in blue with a note explaining why IPv4 and IPv6 can resolve to different locations.
Cloudflare data center markers — Cloudflare data centers now show as orange dots on the map and the one you are connected to is highlighted.
Data center connectors — Each line connects your IP markers to their respective data centers.
Due to the data policies of our geolocation provider, this detailed location is only available for your own IP. Other IP addresses keep the current country-level view.
Extended AS information
The AS card on the IP page now shows additional detail about the network an IP belongs to — including alternate names, the operator website, and an estimate of the AS user population — alongside the AS number and country.
Workflows makes it easier to build reliable multi-step applications that can recover when downstream systems fail. Rollback handlers now receive the original step context via a ctx object for the step being rolled back. This includes ctx.step.name, ctx.step.count, ctx.attempt, and the step config with defaults applied.
The step configuration includes the retry and timeout settings used for that step, so you can customize your step recovery logic according to those fields.
await step.do( "create charge", async () => { const charge = await createCharge(); return { chargeId: charge.id }; }, { rollback: async ({ ctx, output, error }) => { // `output` is the value returned by the step being rolled back. const { chargeId } = output as { chargeId: string }; await refundCharge(chargeId, { // `ctx` is the original step context, including step name, count, attempt, and config. reason: `${ctx.step.name}: ${error.message}`, }); }, rollbackConfig: { // `rollbackConfig` controls retries and timeout for the rollback handler. retries: { limit: 3, delay: "30 seconds", backoff: "linear" }, timeout: "5 minutes", }, },);
Regional Services now supports Regionalized IP Bindings, letting you regionalize traffic at the IP layer for prefixes you bring to Cloudflare through Bring Your Own IP (BYOIP).
Where Regional Hostnames regionalize traffic by hostname, Regionalized IP Bindings let you bind a CIDR from one of your prefixes to a region — ideal for address-map deployments and any service you address by IP rather than hostname. Cloudflare then terminates TLS and processes traffic to those addresses only within the data centers in that region.
Regionalized IP Bindings requires the Regional Services and Regional Services for BYOIP entitlements. Contact your account team to enable them.
Cloudflare Accelerated Mobile Pages (AMP) and Signed Exchanges (SXG) support has reached end of life. The features have been disabled since October 2025, so customers who had them configured should see no change to their traffic.
Customers will no longer be able to configure AMP/SXG through API or rulesets. The Zone API will start throwing errors. Rulesets with the SXG configuration will fail to save until SXG has been removed.
This week's release introduces new managed protection to address a critical pre-authentication OS command injection vulnerability in Ivanti Sentry (CVE-2026-10520).
Key Findings
CVE-2026-10520: An OS command injection vulnerability in Ivanti Sentry allows remote, unauthenticated attackers to execute arbitrary system commands with root privileges. The flaw stems from improper sanitization of input strings parsed during internal configuration handling.
R2 SQL now supports window functions, SELECT DISTINCT, set operations, and additional aggregates, making it easier to write analytical queries without preprocessing your data elsewhere.
Window functions — ROW_NUMBER, RANK, DENSE_RANK, PERCENT_RANK, CUME_DIST, NTILE, LAG, LEAD, FIRST_VALUE, LAST_VALUE, NTH_VALUE, and aggregates with an OVER (...) clause, including PARTITION BY and explicit frames
QUALIFY — filter rows based on a window function result
DISTINCT — SELECT DISTINCT, DISTINCT ON (...), and the DISTINCT modifier on aggregates such as COUNT(DISTINCT ...)
Set operations — UNION, UNION ALL, INTERSECT, and EXCEPT
Grouping extensions — GROUPING SETS, ROLLUP, and CUBE
Exact aggregates — MEDIAN, PERCENTILE_CONT, ARRAY_AGG, and STRING_AGG
Examples
Rank rows with a window function
SELECT customer_id, region, ROW_NUMBER() OVER (PARTITION BY region ORDER BY total_amount DESC) AS rank_in_regionFROM my_namespace.sales_data
Filter with QUALIFY
SELECT customer_id, region, total_amountFROM my_namespace.sales_dataQUALIFY ROW_NUMBER() OVER (PARTITION BY region ORDER BY total_amount DESC) <= 3
Combine tables with a set operation
SELECT customer_id FROM my_namespace.sales_dataEXCEPTSELECT customer_id FROM my_namespace.archived_sales
The named WINDOW clause is not supported — inline the OVER (...) specification at each call site. For the full syntax reference, refer to the SQL reference. For supported features and performance guidance, refer to Limitations and best practices.
The Routes page in the Cloudflare dashboard now shows the routes across all of your connectors — Cloudflare Mesh and Cloudflare Tunnel routes alongside Cloudflare WAN and Magic Transit static routes — in a single table, instead of a separate routes view per product.
From the unified Routes page you can:
Visualize your network with an interactive map that shows how your destinations flow through to your connectors — including equal-cost multi-path (ECMP) routes where the same prefix is served by several connectors. Select a node to filter the table down to the routes behind it.
See every route in one table, with its destination, type, connector, priority, and source, and filter or sort to find what you need.
Create, edit, and delete routes of any supported type without leaving the page. When adding a Cloudflare WAN or Magic Transit static route, you now pick the next hop by connector name instead of typing its IP.
Your existing routes, APIs, and configurations are unchanged — this is a dashboard experience that brings them together in one place. Learn how to add routes and manage virtual networks.
Durable Objects now supports two new location hints for Asia-Pacific: apac-ne (Northeast Asia-Pacific) and apac-se (Southeast Asia-Pacific). Use apac-ne or apac-se when you want finer-grained placement within Asia-Pacific rather than the broader apac hint.
Use the new hints the same way as any other locationHint:
If your users are spread across all of Asia-Pacific, the existing apac hint remains the right choice. Only reach for apac-ne or apac-se when your traffic is clearly concentrated in one sub-region and you want to minimize round-trip time to that audience. The default behavior and what we generally recommended is not adding a location hint unless absolutely needed, this will create the Durable Object as close to the initializing request as possible to reduce latency.
As with all location hints, these are best-effort suggestions. Cloudflare will place the Durable Object in a nearby data center, not necessarily the exact hinted location.
Durable Objects now remain alive for the duration of active outbound connections created via connect() or an outbound WebSocket. Previously, a Durable Object would be evicted after 70-140 seconds of no incoming traffic, even if the object had an open outbound connection, which is a common pattern when streaming responses from a large language model (LLM) over TCP or an outbound WebSocket.
With this change, each active outbound connection prevents eviction. Once all outbound connections close, the standard 70-140 second inactivity window applies before the Durable Object is evicted.
Before: streaming connections were cut off by eviction
After: active outbound connections keep the Durable Object alive
If you are building agents on Cloudflare, this is especially relevant. An agent that streams tokens from an LLM while calling models, or that performs long-running tasks over an outbound connection, now stays alive for the duration of that connection instead of being evicted mid-stream.
Limits:
Each outbound connection keeps the Durable Object alive for a maximum of 15 minutes. After 15 minutes, the connection stops preventing eviction (the connection itself continues operating), and the standard eviction rules resume.
AI agents can now deploy Workers to Cloudflare without first requiring a user to sign up, open a browser-based OAuth flow, click through the dashboard, or create an API token. When an agent tries to deploy without Cloudflare credentials, Wrangler can tell it to rerun with --temporary, then deploy the Worker to a temporary preview account.
To try this with your agent, update to Wrangler 4.102.0 or later, make sure you are logged out (wrangler logout), and then ask your agent to build something and deploy it to Cloudflare. The agent should follow Wrangler's output and deploy using the --temporary flag.
wrangler deploy --temporary
The temporary deployment stays live for 60 minutes. During that window, the agent can verify the Worker, redeploy changes, and return both the live Worker URL and claim URL. Opening the claim URL lets you sign in to or create a Cloudflare account and make the temporary account permanent.
Temporary preview accounts currently support a limited set of products, including Workers, Workers Static Assets, Workers KV, D1, Durable Objects, Hyperdrive, Queues, and SSL/TLS certificates. For supported products, limits, and claim behavior, refer to Claim deployments (temporary accounts).
When you create a new Zero Trust organization, Cloudflare now adds the Cloudflare identity provider as your default login method. Previously, new organizations started with one-time PIN (OTP).
With the Cloudflare identity provider, your users authenticate using their existing Cloudflare account credentials, and authentication is restricted to members of your account. You can still add OTP or connect any third-party identity provider whenever you need to.
This change only applies to newly created accounts. Existing organizations keep the login methods they already have configured. If you would like to use the Cloudflare Identity Provider in an existing account, you must enable it.
exec() is now available for Containers. Use this.ctx.container.exec() to start processes inside a running Container, stream standard input and output, inspect exit codes, and signal each process.
Call exec() from a class extending Container, or from another Durable Object through this.ctx.container. The associated Container must already be running.
This example starts the Container when needed, then reads its Node.js version:
src/index.jsjs
import { Container } from "@cloudflare/containers";export class MyContainer extends Container { async readVersion() { if (!this.ctx.container.running) { await this.start(); } const process = await this.ctx.container.exec(["node", "--version"]); const output = await process.output(); const decoder = new TextDecoder(); return { exitCode: output.exitCode, stdout: decoder.decode(output.stdout), stderr: decoder.decode(output.stderr), }; }}
src/index.tsts
import { Container } from "@cloudflare/containers";export class MyContainer extends Container { async readVersion() { if (!this.ctx.container.running) { await this.start(); } const process = await this.ctx.container.exec(["node", "--version"]); const output = await process.output(); const decoder = new TextDecoder(); return { exitCode: output.exitCode, stdout: decoder.decode(output.stdout), stderr: decoder.decode(output.stderr), }; }}
The command array starts an executable directly, without an implicit shell. Invoke a shell explicitly for pipes, redirects, or variable expansion.
One RPC method can coordinate multiple exec() calls in one caller-to-Durable Object round trip. It can also pass byte-oriented ReadableStream input or return streamed output with flow control.
You can create PlanetScale Postgres and MySQL databases from Cloudflare and bill PlanetScale database usage through your Cloudflare account as a pay-as-you-go customer. Cloudflare contract customers will be able to add PlanetScale usage to their contract in July so reach out to your Cloudflare account team if interested.
Create a PlanetScale database from the Cloudflare dashboard to check out globally distributed Workers optimized for regional data access.
PlanetScale databases created from Cloudflare work with Workers through Hyperdrive. Hyperdrive manages database connection pools and query caching, so you can use PlanetScale as a centralized relational database for Workers applications without changing your database drivers, object-relational mapping (ORM) libraries, or SQL tooling.
PlanetScale usage appears on your Cloudflare invoice each billing period as a dollar total at PlanetScale's standard pricing ↗. You can introspect per-database billing usage via PlanetScale's dashboard ↗.
When you create a PlanetScale database from the Cloudflare dashboard, you receive the same PlanetScale developer experience, including development branches, query insights, and Model Context Protocol (MCP) server support for agents.
Radar has changed how it measures Workers AI model and task popularity.
Previously, popularity was based on the number of unique accounts running inferences against each model or task. It is now based on the number of inferences, giving a more representative view of actual usage volume. This change will affect all new measurements as well as historical data. As a result, the model and task distributions shown on Radar may differ from what you saw previously, and historical trends may shift accordingly.
Cloudflare Fonts now forwards /cf-fonts requests to your origin server when it encounters invalid paths or unexpected runtime errors, instead of returning 4xx or 5xx responses directly. This update also adds additional input validation to enhance security.
