October 2026 Summaries
22 posts from Upstash
Filter
Month:
Year:
Post Summaries
Back to Blog
QStash workflows now support step-specific flow control and retry settings through a third configuration argument to `context.run`, allowing individual operations to receive different treatment without affecting the rest of a workflow. This is useful for actions such as payment API calls that have strict concurrency limits or unreliable external services that need additional retry attempts. A step-level flow-control configuration, including a shared key and parallelism limit, overrides the workflow run’s flow-control settings only for that step and can limit concurrency across all runs using the same key. Developers can also set per-step retry counts and delays, with unspecified options inheriting the workflow’s defaults, and can combine retry and flow-control settings together. QStash workflow logs display each step’s effective configuration, including flow-control and retry details, in the Config tab.
Oct 28, 2026
600 words in the original blog post.
Context7’s comparison of its Search API with Exa 3.0 Lite and Auto on Exa’s 307-question WebCode coding-retrieval benchmark found nearly identical correct-answer rates, at 85.7% for Context7 and 86.2% for both Exa options, using the same answer and judge models across three production runs. Context7 reported substantially fewer incorrect answers, 2.3% versus 7.3% for Lite and 8.0% for Auto, while returning shorter responses of roughly 1,050 tokens per query compared with about 4,650–5,020 tokens for Exa, although Exa searches were faster, with median latencies of 0.6 seconds for Lite and 1.3 seconds for Auto versus 3.7 seconds for Context7. The post attributes Context7’s lower error rate to more concentrated results that more often contain the relevant answer and cause the answer model to decline rather than guess when evidence is absent. It also notes that the services succeeded on different subsets of questions, illustrating that their retrieval strengths vary, and provides public API access and pricing information for Context7 alongside Exa’s per-search costs.
Oct 27, 2026
1,334 words in the original blog post.
A Customer Success leader at Upstash describes building a persistent personal work assistant around Hermes, an open-source agent from Nous Research hosted on an always-on Upstash Box and accessed through Slack on mobile and Hermes Desktop on a computer. Connected to Gmail, Slack, Calendar, Drive, support tools, and cloud browsing through Upsy, the agent prepares meeting briefs, triages email, monitors customer signals, identifies upgrade or churn risks, drafts outreach, and tracks ideas, while requiring human approval for consequential actions such as sending emails or changing billing. The author values Hermes’s reusable, self-improving skills, model flexibility, open-source transparency, and ability to use temporary cloud machines for larger tasks while keeping its primary environment lightweight. The setup includes scheduled Slack briefs, documented operational rules, a sandboxed Upstash account for temporary workloads, and a projected second agent for marketing content, positioning the system as a comparatively low-cost, continuously available alternative to consumer chat assistants such as Grok Bot.
Oct 26, 2026
1,832 words in the original blog post.
Upsy Meet, a meeting room for AI agents, uses the MCP Tasks extension to make agent listening more efficient by replacing repeated long-polling calls to its `meeting_wait` tool with asynchronous tasks that Claude Code can monitor while continuing other work. Previously, agents repeatedly held server requests open for up to 50 seconds, consuming model turns, function time, and preventing concurrent work; with Tasks support in newer Claude Code versions, the server returns a task immediately and the client polls it at server-directed intervals until new messages arrive. Upsy stores only minimal wait state in Redis rather than launching background jobs, since each task merely checks whether messages have appeared since a given point in an existing meeting stream. The implementation prevents duplicate waits per participant, ignores an agent’s own messages, returns existing messages immediately, gradually lengthens polling intervals during inactivity, and routes task polling through a lightweight authenticated request path. Clients without the Tasks extension continue using conventional waiting, while an SDK issue in TypeScript SDK version 2.1.0 required upgrading to version 2.3.1 for `tasks/get` and cancellation handlers to work correctly.
Oct 24, 2026
1,415 words in the original blog post.
Claude Code now supports MCP Tasks and MCP Elicitations, enabling tools to run long operations asynchronously and request user input during execution. Tasks, introduced in the July 2026 MCP release and available in public beta in Claude Code CLI version 2.1.304, let compatible servers immediately return a task ID while Claude Code polls for completion, continues other work, and can cancel tasks; unsupported clients instead receive conventional polling and cancellation tools. Elicitations present terminal forms for confirmations or other structured responses, allowing tools or workflows to pause until users respond, though task-based questions require an interactive terminal session. The @upstash/mcp-toolkit/tasks package provides task support for MCP servers across frameworks, with in-memory operation for simple deployments and Redis plus QStash or Upstash Workflow options for serverless, multi-instance, retryable execution. QStash runs each task in a separate function invocation and retries failures with backoff, while WorkflowDispatcher can split longer work into steps and wait for user responses without occupying a function.
Oct 23, 2026
1,588 words in the original blog post.
MCP Apps are interactive HTML interfaces delivered by Model Context Protocol servers within AI conversations, allowing users to work with data through tables, charts, filters, and controls instead of relying solely on repeated text prompts. Introduced as the first official MCP extension in January 2026, they combine a tool with a UI resource, which compatible hosts such as ChatGPT, Claude, and VS Code render in a sandboxed iframe while mediating all tool calls and user confirmations. Upstash uses this capability to embed much of its console in supported chat clients, enabling management of Redis, QStash, Workflow, Vector, Search, Box, and Blob resources, including browsing and editing Redis keys, publishing messages, viewing logs, and administering databases. The interface reuses the user’s authorized MCP connection rather than storing credentials, honors read-only permissions, keeps database tokens away from the model, and routes sensitive operations through server-side tool calls. Its Redis Insights features also let agents examine traffic, latency, performance, and limits to help investigate issues such as database slowdowns, while clients without MCP App support can still use the server’s standard tools.
Oct 21, 2026
977 words in the original blog post.
Upstash Box provides isolated Docker-based sandboxes, called boxes, where AI agents can safely run generated code, manage files, install packages, access a shell, and optionally use network connectivity without affecting a user’s local machine or other users’ environments. The guide explains how these sandboxes support applications such as executing LLM-written scripts, persistent per-user agent servers, coding agents that modify repositories, parallel tasks, browser automation, and model benchmarking, while emphasizing that credentials and network access must be restricted because sandbox isolation cannot reverse external API calls or leaked secrets. It demonstrates creating a free account, configuring the Node.js or Python SDK, launching boxes, executing commands, handling failed exit codes, writing and reading files, setting timeouts for generated code, and running coding agents such as Claude Code. Boxes can be paused to save compute while retaining files, snapshotted and restored to reuse environments, or deleted for cleanup, with ephemeral boxes available for short-lived command and file tasks. The free tier includes up to 10 concurrent boxes, 5 monthly CPU hours, 2 vCPUs, 4 GB memory, 5 GB storage per box, and a limited built-in agent-token budget, while paid plans offer higher capacity, longer or no idle shutdowns, and either usage-based or fixed monthly pricing for continuously available services.
Oct 19, 2026
2,538 words in the original blog post.
Upstash Box can be used to build a Lovable-style application builder by assigning each project its own persistent box containing a coding agent, source code, Git history, and a running development server with a public preview URL. This design avoids synchronizing generated files between a central server and sandbox, while snapshots provide reusable templates and efficient project forks, Git commits support versioning and restores, and idle boxes pause to reduce compute costs but can resume through the preview URL in roughly two seconds. A Next.js application manages project metadata, chat history, locks, and rate limits through Redis while forwarding prompts to agents such as Claude Code, Codex, or OpenCode running inside each box. Tests using a small 2-vCPU, 4-GB box and Claude Sonnet 5 showed project creation in 8 to 15 seconds, prompt execution in 19 to 49 seconds, restores in about 3 seconds, and low CPU charges because model tokens, rather than sandbox compute, are the primary expense.
Oct 15, 2026
1,572 words in the original blog post.
Upstash Box is presented as a container-based workflow for running multiple coding agents concurrently without the file, process, port, network, and credential-sharing conflicts that arise when agents use one repository checkout or Git worktrees on the same machine. Users prepare a seed box by cloning and optionally installing a repository, snapshot it, launch one isolated box per task, run agents and tests in parallel, preserve changes as patches or pull requests, and delete the temporary environments afterward. The approach can also run several attempts for a difficult task and select a result using independent acceptance tests, test modifications, and diff size as a tiebreaker. Isolation does not eliminate merge conflicts among independently generated changes, and agents can still access environment variables and make outbound requests unless network policies limit them. Pricing combines active CPU use with model-token costs, making Box relatively inexpensive for agents that spend much of their time waiting for models, while alternatives such as E2B, Daytona, or Modal may be better for needs including stronger VM isolation, running-process snapshots, GPUs, custom cloud deployment, Python-focused workloads, or regional data requirements.
Oct 15, 2026
2,308 words in the original blog post.
Cloud computers expand AI agents beyond API access by providing isolated environments with shells, persistent filesystems, browsers, Git, network controls, and lifecycle management, allowing them to perform coding, browser-based, and long-running personal-assistant tasks. The comparison presents three main approaches: self-managed virtual machines, which offer control but require users to handle security and provisioning; desktop sandboxes such as E2B Desktop and Daytona Computer Use, which support GUI applications through VNC; and agent-oriented cloud computers such as Upstash Box, which provide persistent workspaces, built-in coding agents, headless browsers, snapshots, scheduling, and API-driven provisioning. Upstash Box is positioned for per-user and coding-agent workflows, with an example lifecycle covering repository setup, automated code changes and testing, browser validation, snapshots, automatic pausing, scheduled work, and always-on assistants. Its stated limitations include container-based rather than VM isolation, AWS us-east-1 availability, no GPU support, and no custom Docker images yet, while pricing emphasizes active CPU usage, low-cost paused storage, and a flat keep-alive option for continuously running agents.
Oct 14, 2026
2,181 words in the original blog post.
Backblaze B2 charges $6.95 per TB monthly for storage, with the first 10 GB free, free uploads and most API calls, and free downloads capped at three times average stored data before charging $0.01 per GB; as a result, heavily accessed application files can make egress costs dominate storage costs. The comparison argues that Upstash Blob may be less expensive for user-facing images, PDFs, and exports because it includes 1 TB of monthly downloads regardless of storage volume, though it has higher storage, read, upload, and post-allowance download rates. B2 can avoid download charges through partner CDNs such as Cloudflare, but configuring public or private access requires domain, rewrite, caching, and sometimes request-signing setup, with misconfiguration potentially exposing other public buckets through a shared host. Both services support direct browser uploads, but B2 relies on S3-compatible presigned URLs and developer-managed CORS, progress, retries, and multipart behavior, whereas Upstash Blob provides SDK tools for upload policies, signed URLs, multipart transfers, and client-side progress handling. Overall, B2 is presented as suited to inexpensive backups and archives that are rarely retrieved, while Upstash Blob is positioned for globally delivered application content with frequent downloads.
Oct 12, 2026
1,649 words in the original blog post.
A study of nine verified dangerous documentation pages examined how six retrieval tools and eight language models handle instructions that could lead coding agents to execute unsafe actions, such as downloading unpinned code with root privileges or exposing credentials. Five conventional search tools returned large portions of page content and included harmful passages in 23% to 42% of searches, while Context7, which screens libraries and supplies shorter matched excerpts, did so in 10% of searches; after excluding cases where risky parameters were inherently part of legitimate API answers, Context7 returned none compared with 24% for web search. Most tested pages were not deliberate attacks but good-faith documentation that nevertheless directed users to give durable access to untrusted or unverifiable third parties, highlighting risks beyond prompt injection. When models received web-search results containing dangerous instructions, they repeated them as recommended steps in 35% of applicable answers, usually without warnings, whereas Context7 led to such repetition in under 1% of answers. The authors caution that the small, nonrandom sample cannot measure web-wide prevalence or assess specific coding assistants, but conclude that retrieval-tool design strongly influences what unsafe material reaches an agent.
Oct 11, 2026
1,806 words in the original blog post.
Amazon S3 pricing combines storage, API requests, data transfer, and optional features, with S3 Standard in US East priced at $0.023 per GB-month, $0.005 per 1,000 write requests, $0.0004 per 1,000 reads, and $0.09 per GB of internet egress after the first 100 GB monthly allowance. Lower-cost storage classes such as Standard-IA and Glacier reduce storage rates but can add retrieval fees, minimum object sizes, minimum retention periods, and archive metadata charges, making them less economical for small or short-lived files. Request and transfer charges can exceed storage costs substantially for applications that handle many small objects or serve significant download traffic, while versioning, incomplete multipart uploads, lifecycle transitions, Intelligent-Tiering monitoring, and reporting can create additional fees. CloudFront can reduce or restructure delivery costs through free allowances and flat-rate plans, although S3 still charges for storage and cache-miss requests. The article recommends calculating bills by adding storage, request, and egress costs using AWS’s published regional price lists, illustrates that transfer dominates example workloads, and compares S3 with Upstash Blob, which offers simpler global pricing, an included CDN, and lower transfer rates but is less suited to low-cost archival storage or workloads tightly integrated with AWS services.
Oct 09, 2026
3,695 words in the original blog post.
Personal AI assistants are increasingly divided between hosted services, represented by Grok Bot and Meta Muse, and self-hosted open-source options, OpenClaw and Hermes Agent. Hosted products provide the simplest setup because their companies operate the underlying computers, manage maintenance, and select available models, but users have less control over model choice, data storage, and connector availability. Grok Bot emphasizes persistent cloud computers, multiple collaborating bots, workflow skills, and broad MCP support, while Muse focuses on consumer errands such as email, bookings, and shopping, with transaction safeguards and limited availability in the United States and Canada. OpenClaw and Hermes allow users to select model providers, control stored agent data, connect through numerous messaging apps, use MCP servers, and schedule automations, but require continuous hosting, updates, security management, and careful handling of third-party skills or secrets. OpenClaw has a larger ecosystem and supports many chat platforms, whereas Hermes highlights flexible command execution, model switching, and learning repeated routines as skills. The choice depends largely on whether a user prioritizes convenience, everyday task assistance, autonomous computer work, ownership and customization, or always-on self-managed operation, with the text recommending gradual permission grants because all of these agents may be able to send messages, make purchases, or delete information when authorized.
Oct 09, 2026
1,800 words in the original blog post.
MCP agents face limitations with long-running work and reacting to external changes, which the emerging MCP Tasks extension and proposed MCP Events model aim to address. Tasks allow servers to return a handle immediately for background work that clients later poll for completion, while Events use signed webhooks to notify agents when subscribed real-world conditions occur; as of October 2026, native client support remains limited, with ChatGPT supporting Events in Work chats but not native Tasks. Upstash’s `@upstash/mcp-toolkit`, built atop the official TypeScript SDK, provides portable task and event layers that work through standard MCP tools and endpoints, using Redis for persistent task and subscription state and QStash or Workflow for durable execution and delivery. The toolkit emphasizes authentication, per-user ownership, authorization before queuing tasks and delivering events, webhook verification, signing, retries, and subscription expiry. Tests with ChatGPT showed that long-running tasks currently operate through ordinary tool calls and polling, whereas event subscriptions can trigger ChatGPT notifications, though observed delivery-to-notification latency of roughly three minutes may not suit time-critical uses.
Oct 09, 2026
2,156 words in the original blog post.
The post describes using Grok Bot with the Upstash MCP server and Upstash Box to give an AI assistant access to a persistent, sandboxed cloud computer capable of running shell commands, managing files, cloning GitHub repositories, opening pull requests, hosting public previews, taking screenshots, and scheduling tasks. After OAuth-based setup, the author tested the workflow by asking the bot to build and host a Snake game, then expanded it to inspect repositories, run builds and tests, diagnose code, and implement fixes. The system can combine connections to tools such as Slack, Linear, GitHub, Box, and Blob so that a bug report can be turned into a ticket, code change, tested branch, and screenshot-equipped pull request, while GitHub authentication is managed through an app rather than exposing tokens to the agent. It also supports shared Box environments in which multiple agents can implement and review changes on the same files and branch. The author emphasizes voice-driven use from a phone, allowing tasks to be delegated remotely, and recommends starting with a small project that produces a preview URL, linking GitHub first, naming persistent boxes clearly, using supplied remote-work skills, and scheduling recurring automated checks.
Oct 07, 2026
1,536 words in the original blog post.
A plain React SPA can implement secure, direct-to-storage file uploads by combining a drag-and-drop library such as react-dropzone, a small Hono server for authentication and signed upload authorization, and object storage such as Upstash Blob. Storage bearer tokens must remain on the server because Vite exposes client-prefixed environment variables in browser bundles, while signed URLs restrict uploads to a specific object, method, and short time period so files can bypass server body-size limits. The example uses an Upstash Blob upload handler to enforce image and size constraints, create unique storage paths, authenticate users, and save completed-upload metadata to Redis idempotently, with Hono routes exposing upload and album-listing endpoints. React hooks provide per-file upload status, progress, cancellation, retrying, and previews, while client-side checks improve usability but server-side validation remains necessary. Multipart uploads support large files, retry temporary network failures, renew expired signatures, and can resume after reloads in some cases, but applications must separately clean up abandoned multipart uploads, handle repeated completion callbacks safely, and treat uploaded content as potentially unsafe. Upstash Blob and Cloudflare R2 are presented as economical for lower-volume use, while R2 may suit high-egress workloads, UploadThing offers a more packaged widget experience, Cloudinary supports media transformations, and Amazon S3 fits applications already based on AWS.
Oct 07, 2026
2,910 words in the original blog post.
Daytona is presented as a fast, polyglot environment for running AI-generated code, offering container-based sandboxes, optional VM classes, GPUs, and Windows and macOS support, but teams may seek alternatives because of wall-clock billing, lower-tier network restrictions, a lack of built-in agent orchestration, default idle shutdowns, and the archiving of its open-source repository after core development became private in June 2026. Upstash Box is positioned for managed coding-agent workflows through built-in agent execution, browser testing, typed results, persistent files, and active-CPU billing, while E2B and Vercel Sandbox emphasize Firecracker microVM isolation, with E2B also providing stateful code interpreters and Vercel targeting users already on its platform. Modal is recommended for Python-centric GPU workloads, Tensorlake for VM snapshots and bring-your-own-cloud deployments, and Northflank for lower-cost compute, GPUs, and running workloads in a customer’s cloud account. Organizations needing maintained self-hosted options may consider E2B Embed or Beam’s Beta9, whereas Daytona may remain suitable for users requiring GPUs, desktop operating system sandboxes, Go, Ruby, or Java SDKs, VNC-based computer use, or EU availability.
Oct 06, 2026
1,893 words in the original blog post.
Selecting cloud storage for a side project depends mainly on free-tier limits, egress charges, setup complexity, serverless compatibility, and browser-upload support rather than minor storage-price differences. Upstash Blob is presented as a strong default for most Next.js and serverless projects because it uses a single bucket token, offers 1 GB of storage, a hard free-tier limit that prevents unexpected charges, and up to 1 TB of free monthly egress on paid usage, while running on Cloudflare R2 infrastructure and retaining S3 compatibility. Cloudflare R2 provides the largest free storage allowance at 10 GB and free egress but requires a more S3-like setup, whereas AWS S3 suits projects already integrated with AWS despite more involved IAM configuration and comparatively high egress costs. Supabase Storage is positioned for existing Supabase users, Cloudinary for image and video transformations, and other services such as Vercel Blob, Bunny Storage, and UploadThing offer different pricing or platform trade-offs. The example implementation shows how an Upstash Blob bucket can store JSON backups, list and prune old files, restore the newest version, and create temporary signed download links, while its browser-upload mechanisms allow large files to bypass serverless request-size limits through upload handlers or signed upload URLs.
Oct 05, 2026
2,611 words in the original blog post.
Dub replaced its Jev-based semantic emoji-search fallback with a single Upstash Vector index to let users find emoji by meaning while reducing recurring input-token costs. Jev required sending and scoring the full catalog of 1,941 emoji on every fallback search, forcing the catalog to be chunked and making costs scale with both catalog size and search volume. The vector approach embeds each emoji’s label and tags once during a seed process using data from emojibase, stores the resulting vectors with emoji metadata, and embeds only the user’s short query at search time to retrieve the closest matches. Searches return up to 18 results above a 0.45 similarity threshold, while exact name matching remains the first step; the semantic route is session-protected, rate-limited, debounced on the client, and cached in memory. Because the emoji catalog is small and rarely changes, it fits within Upstash Vector’s free tier of 10,000 daily requests, with pay-as-you-go pricing estimated at about $4 for one million searches.
Oct 05, 2026
916 words in the original blog post.
A benchmark of 44 API changes across 32 JavaScript and Python libraries found that Context7 returned responses containing only current APIs in 59% of searches, compared with 25% for Exa fast search and 18% for Exa’s AI SDK tool, while using roughly one-tenth as many tokens; Exa was faster, however, particularly in fast-search mode. The test used version-specific questions based on documented 2026 API renames or removals, ran each question twice through all tools, and excluded cases where official current documentation still showed obsolete examples. Exa’s stale results commonly came from GitHub repositories, Stack Overflow, archived projects, old guides, and unofficial mirrors, though Context7 also sometimes surfaced outdated versioned documentation or first-party source files that lagged behind current docs. Context7 is presented as most useful for official library documentation, whereas Exa’s wider web coverage is better suited to issues, announcements, and news. Context7 maintains freshness by indexing owner-published sources, prioritizing current versions, regularly refreshing popular libraries, supporting owner-triggered updates, and allowing library and version hints in searches. The page also demonstrates integrating Context7 as a documentation-search tool in an AI SDK coding agent and notes planned improvements in freshness detection, version matching, and result relevance.
Oct 02, 2026
2,432 words in the original blog post.
Upstash AgentKit for TanStack AI provides Redis-backed production state for TanStack AI agents, addressing serverless environments where in-memory state is lost between requests or instances. Its modular adapters integrate with existing TanStack AI middleware to support persistent chat transcripts, run records and approvals; resumable streaming responses stored in Redis Streams; distributed locks for operations that must run only once; and long-term, user-scoped memory retrieved through Redis Search without embeddings. The package also offers configurable tool-result caching for safe, repeatable calls, rate limiting before model execution, and RAG search, aggregation, and count tools over application documents with typo-tolerant querying. Configuration relies on Upstash Redis environment variables, and developers can independently combine the features they need, keeping conversations, streams, locks, memories, cached results, and search indexes in a single Redis database.
Oct 01, 2026
1,315 words in the original blog post.