MCP Drops Protocol-Level Sessions, Enabling Stateless Horizontal Scaling

·1 views

The Model Context Protocol (MCP) has updated its specification to move away from protocol-level session management toward a stateless request model. Previously, Streamable HTTP connections relied on session identifiers that tied clients to specific server instances, complicating horizontal scaling across multiple nodes. The new approach requires each request to carry enough context to be processed independently, allowing load balancers to route traffic to any available server instance without sticky sessions or shared session storage. A developer documented migrating a currency converter MCP server to the updated stateless model using the new split TypeScript SDK package, specifically @modelcontextprotocol/server. The change is intended to make MCP servers behave more like standard cloud-native HTTP services, reducing deployment complexity while still permitting application-level state where genuinely needed.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Builder Learns Positive Feedback Does Not Equal Product Demand

A developer built Rizzzler, an open-source link-in-bio platform with features like community chat, coins, and milestones, aiming to offer a more interactive alternative to existing tools. After submitting the product to platforms like Product Hunt and SaaSFrame, it received an initial traffic spike and positive comments praising its design and interface. However, signups did not sustain, and traffic quickly dropped off once the launch momentum faded. When the developer spoke directly with potential users, a recurring response emerged: the product was well-made, but people simply did not feel they needed it. The experience highlighted a key lesson — that user approval and genuine demand are distinct, and adding features cannot compensate for an unproven need.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why 'Delete' in AI Apps Often Leaves User Data Scattered Across Backend Systems

Deleting a user's data in AI-powered applications is far more complex than removing a visible chat entry, as a single agent interaction can spread data across prompts, memory stores, vector indexes, tool logs, and analytics systems. Developers writing for DEV Community warn that many apps only delete the UI-facing record while leaving sensitive data intact in multiple backend layers. This creates a trust and compliance risk, since deleted content can resurface indirectly through memory summaries, cached prompts, embeddings, or support traces. The guide recommends that AI app builders first map every storage surface where user data can land before writing any deletion logic. Proper deletion pipelines must address each surface with tailored actions — such as hard deletes, payload redaction, or pseudonymization — rather than treating all records the same way.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Host a Free Website Using GitLab Pages and CI/CD Pipelines

GitLab Pages allows developers to host HTML and CSS websites for free directly from a GitLab repository. The process requires creating a .gitlab-ci.yml file that instructs GitLab on how to deploy the site, after which a CI/CD pipeline publishes it at a unique URL. A successful deployment depends on a valid YAML configuration and a correctly structured public folder containing index.html at its root. Common pipeline failures include invalid YAML syntax, incorrect indentation, or attempting to create a folder that already exists. For managing multiple distinct projects, GitLab recommends maintaining a separate repository for each website to avoid conflicts and simplify maintenance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer quarantines AI-written code for three days despite all 82 tests passing

A software developer discovered an uncommitted code change in their open-source repository that had been written by a parallel AI coding session working on an unrelated task. The AI had independently identified and fixed a bug in a browser automation tool, verified the fix, and left the improved code in the working tree without committing or notifying anyone. Despite all 82 tests passing, the developer chose to quarantine the change for three days rather than commit it immediately, citing concerns about missing context and the limits of test coverage. Upon careful review after the deadline, the developer found a subtle but meaningful documentation error: the new code incorrectly described coordinate values as page-level when they were actually relative to each frame's viewport. The bug was ultimately fixed by correcting the comments and clarifying that callers must account for frame offsets using a field the AI had already included in the diff.

0 comments Read more at DEV Community