Developer Platform Changelog

Sent emails can now be previewed directly from the Email Service Activity log, showing the rendered HTML, text, headers, attachments, and raw RFC 5322 source. Turn on Email preview in sending domain settings; new domains after 2026-07-02 have it enabled by default.

Wrangler now includes wrangler flagship, a command suite for creating Flagship apps, managing feature flags, and controlling release exposure — including rollout, split, and targeting rules — from the terminal and CI/CD pipelines.

Email Sending events can now be received through Queues event subscriptions, with six lifecycle event types including delivery, bounce, and complaint messages. Each subscription is scoped to one sending domain.

The legacy Workers KV API routes under /accounts/{account_id}/workers/namespaces/* are deprecated with an October 15, 2026 sunset date. Migrate to the replacement routes under /accounts/{account_id}/storage/kv/namespaces/ via a direct URL path substitution — request parameters and response payloads are identical.

R2 Data Catalog now rewrites and clusters manifest files by partition as a pre-step to compaction, reducing metadata I/O overhead for query engines. The optimization is automatic for tables with compaction enabled.

R2 Data Catalog now accepts read-only API tokens, allowing query engines to read data without requiring a read-write token. This follows the principle of least privilege for read-only workloads.

Agents connected to MCP servers via addMcpServer can now handle elicitation requests, allowing servers to request user input — form mode for structured data or URL mode for consent before opening an out-of-band flow. Handlers are configured per mode in onStart(), and the SDK stores advertised modes to survive Durable Object hibernation.

The Markdown Conversion service now supports an output.format option set to "text" to receive plain text with Markdown syntax removed. The default remains "markdown", so existing conversions are unchanged.

Step retries now accept a function passed to retries.delay that calculates the next delay based on the failed attempt and thrown error, allowing retries to wait longer after rate-limit errors but retry sooner after transient failures, and to parse downstream API guidance like Retry-After headers.

Wrangler automatic resource provisioning now supports Flagship bindings. Adding a flagship binding without an app_id lets Wrangler connect an existing Flagship app or create a new one during deployment, writing the app_id back to the configuration file.

Accounts without existing key-value backed Durable Object namespaces can no longer create them; new namespaces must use the SQLite storage backend. Deployments using new_classes migrations fail with an error for affected accounts.

Wrangler now collects npm package dependency information from package.json during wrangler deploy and wrangler versions upload, including each dependency's name, declared version range, and exact installed version. This data enables dependency analytics and future supply chain security features such as vulnerability alerting. You can opt out by setting dependencies_instrumentation.enabled to false in your Wrangler configuration.

Moondream 3.1, a fast vision language model with 9B total parameters and 2B active, is now available on Workers AI. The model supports querying, captioning, point detection, and object detection with a 32K token context window, and streams first tokens in 20–30 ms for real-time edge vision workloads.

Workers AI's toMarkdown feature now converts GIF and BMP image files to Markdown in addition to JPEG, PNG, WebP, and SVG. AI Search automatically includes GIF and BMP files when processing ingested content on the next index sync, with no configuration changes required.

The AI Search list items endpoint now accepts a key query parameter to look up a single item by its exact object key without paging through the full list, complementing the existing item_id filter. Keys can be combined with the source parameter to disambiguate when the same key exists across multiple data sources.

Cloudflare Drop lets you upload a folder or ZIP file of static assets and get a temporary live preview that stays live for 1 hour, shareable immediately without creating an account. After testing, you can claim the deployment to sign in or create a Cloudflare account and keep the site live permanently, then add a domain, enable observability, or control access.

Workflows pricing now includes per-step billing on Workers Paid plans, effective no earlier than August 10, 2026. Paid plans include 500,000 steps per month with overage charges of $0.80 per 100,000 steps; Free plans get 3,000 steps per day with no charges beyond included amounts. Storage pricing remains unchanged at $0.20 per additional GB-month above the 1 GB-month included.

Browser Run now supports a standalone /accessibilityTree endpoint that returns the browser's structured view of a rendered page—roles, names, states, values, and hierarchy—enabling AI agents and automation workflows to access page structure directly without parsing HTML or processing screenshots. The endpoint supports interestingOnly and root parameters for filtering semantically meaningful nodes or capturing subtrees.