December 2025 Summaries

Modern application integrations commonly rely on OAuth and OpenID Connect for client authentication, with client_secret being the default method due to its simplicity and ease of use. This approach, akin to using a password, involves a shared secret between the client and the authorization server, which poses certain risks such as exposure and the need for careful rotation. Alternatively, private_key_jwt offers a more secure method by using asymmetric cryptography, where a client signs a JWT with a private key, and the server verifies it with a corresponding public key. This method reduces the risk of impersonation, provides short-lived authentication artifacts, offers stronger attribution for requests, and simplifies credential management. While client_secret remains suitable for low-risk, controlled environments or when rapid development is prioritized, private_key_jwt is recommended for enhanced security and long-term reliability, addressing vulnerabilities inherent in shared secret models.

The text provides a comprehensive tutorial on building a confidential AI agent that interacts with MCP servers using OAuth tokens, focusing on the implementation of Client ID Metadata Documents (CIMD) for client identity. It details the process of constructing and hosting a CIMD JSON document and a JWKS endpoint to facilitate secure and scalable OAuth authentication across MCP servers. The tutorial explains the end-to-end flow, starting from client registration, generating and hosting a CIMD document, implementing the OAuth Authorization Code flow with PKCE, and exchanging authorization codes for tokens using a signed client assertion via private_key_jwt. Additionally, it covers the server-side validation processes, including checking JWT claims and verifying the client’s identity, ensuring secure communication. The guide includes Python code snippets for generating key pairs, building JWTs, and conducting token exchanges, emphasizing the importance of exact URL matching and JWT validation in maintaining security and functionality.

At MCP Night: The Holiday Special, Reilly Wood from Datadog demonstrated why SQL is a superior interface for AI agents handling large-scale data compared to traditional freeform commands or proprietary query languages. In a side-by-side demo, agents using SQL queries completed tasks more efficiently than those using Datadog's traditional tools, highlighting SQL's ability to manage context and perform efficient data aggregation. At Datadog's scale, SQL's structured nature not only enhances performance but also simplifies the AI agent's workflow by reducing the need for managing intermediate states or context windows. This approach aligns with broader industry trends, as both Datadog and Supabase recognize SQL's precision and reliability in modeling AI queries on internal systems, offering a structured, scalable solution for automated incident investigation and data analysis. The event underscored the idea that SQL, despite its constraints, provides the necessary precision for production systems, suggesting a clear preference for SQL over bash-like scripting in observability platforms.

At MCP Night: The Holiday Special, Craig Cannon from Supabase turned a 90s Christmas movie, "Jingle All The Way," into a live demo showcasing the potential of SQL as an interface for AI agents. The demonstration featured the Turbo-Man Tracker, a real-time inventory system using a Supabase database, where Cannon utilized natural language through the Supabase MCP server to generate SQL queries instantly. This approach highlighted how natural language processing can make database interactions more accessible by eliminating the need for direct SQL knowledge, allowing users to describe their data needs while the system generates and executes the appropriate queries. The event underscored SQL's role as a universal interface for AI, combining precision, flexibility, and familiarity, making data access frictionless for developers. The playful and memorable demo emphasized the importance of engaging presentations in driving protocol adoption, leaving a lasting impression on the audience about the potential of MCP and natural language interfaces.

At MCP Night: The Holiday Special, Den Delimarsky presented a breakthrough in MCP authentication, showcasing a new simplified process that eliminates the need for complex pre-configuration and Dynamic Client Registration, which were previously required for connecting to authenticated services. This advancement, part of an updated MCP specification, introduces Protected Resource Metadata and a novel client identity model, allowing developers to authenticate without being OAuth experts. By providing a metadata endpoint, developers can let clients manage authentication, significantly reducing the complexity and time needed to connect MCP servers to services like GitHub and Google Calendar. The demonstration highlighted how this innovation resolves major adoption barriers for MCP by simplifying the authentication process, thereby enhancing the practical utility of AI agents through easier integration with essential services. This change represents a significant stride forward in MCP's usability and accessibility, offering a more straightforward path for developers to implement authentication in their servers.

At MCP Night: The Holiday Special, WorkOS CEO Michael Grinich highlighted the rapid evolution of the Model Context Protocol (MCP) from a local file system tool to an industry-standard protocol now housed under the Linux Foundation. Originally a way for local environments like Macs to interface with AI models, MCP's potential led to the development of remote servers, enabling connections with services such as GitHub and Salesforce, and accelerating its adoption among major tech companies like OpenAI and Google. The event celebrated MCP's first anniversary, noting the protocol's explosive growth and recent security challenges, which are now a primary focus for the MCP steering committee. Key updates to the protocol include simplified authentication through Protected Resource Metadata, support for asynchronous operations, and enhanced orchestration via bidirectional communication. The announcement of the Agentic AI Foundation—launched with contributions from companies like Anthropic, Block, and Google—marked a significant milestone, providing vendor-neutral governance and addressing enterprise adoption concerns. The keynote underscored MCP's role in the AI era as a universal connector for large language models, positioning it alongside foundational protocols like TCP and HTTP, and setting the stage for its transition to a community-governed standard.

At MCP Night: The Holiday Special, Paul Irish showcased the integration of Chrome DevTools' MCP server, which facilitates programmatic debugging of browser sessions by AI agents, starting with extensive performance traces of around 15 million lines. This innovation addresses the challenge of unevenly distributed expertise in performance debugging by enabling AI to perform initial analyses, thus highlighting critical insights from complex data. The MCP server distills large traces into concise summaries, allowing agents like Gemini CLI or Claude Code to access and utilize browser session information directly, bridging the gap between context and coding agents. By embedding the Chrome team's expertise into these tools, developers can leverage AI-assisted analysis for performance issues, making it accessible even without specialized knowledge. Paul Irish also pointed to potential future developments in standardizing performance data handling, which could enhance the integration across various development tools. The demonstration underscored the practical application of AI in streamlining performance debugging, making the DevTools-to-agent pipeline a current reality rather than a theoretical concept.

At MCP Night: The Holiday Special, Cloudflare's VP of AI & Developers, Rita Kozlov, introduced "Code Mode," a novel approach that could reshape the future of Multi-Cloud Platform (MCP) operations by enabling AI agents to generate and execute code instead of relying on direct tool calls. This method showcased significant advantages in efficiency, particularly in complex tasks, by reducing token consumption—a vital consideration for enterprise-level applications. While traditional tool-calling requires multiple interactions, Code Mode allows for code generation that can efficiently manage tasks through loops and conditionals, as illustrated in a demo where both methods created 31 events, but Code Mode consumed 81% fewer tokens. This approach leverages Cloudflare's Workers Loader for instant sandbox execution, highlighting a potential hybrid model that combines direct tool invocation for simple tasks and code generation for more complex operations, pointing to a more efficient future for MCP applications. Cloudflare is currently expanding access to Code Mode, which is in closed beta, and is seeking a product manager to further develop their AI and MCP strategies.

In 2025, role-based access control (RBAC) has become essential for B2B SaaS platforms, with mature authorization systems now a requirement for securing enterprise deals and ensuring compliance. Rather than building custom permission systems, SaaS teams can leverage a range of RBAC providers offering prebuilt infrastructure to streamline development and standardize permission logic. Key considerations when selecting an RBAC provider include multi-tenant awareness, customizable roles, enterprise-ready integrations, developer experience, operational overhead, and cost transparency. Among the top RBAC solutions, WorkOS is highlighted for its intuitive, developer-friendly system tailored for B2B SaaS, while other providers like Permit.io, Auth0, Logto, and Zitadel offer varying strengths, particularly for niche requirements or open-source preferences. Choosing the right RBAC provider can significantly impact a product's enterprise experience, influencing everything from user onboarding to procurement processes.

On December 10, 2025, MCP Night, The Holiday Special transformed the Regency Ballroom in San Francisco into a pivotal event for the AI ecosystem, marking the public unveiling of the Agentic AI Foundation, which now serves as the new home for the Model Context Protocol (MCP) under the Linux Foundation with backing from major companies such as Anthropic, OpenAI, Block, Microsoft, and Google. The event highlighted MCP's evolution from a project by Anthropic into a vendor-neutral, community-governed standard, with new protocol updates simplifying authentication and enabling asynchronous operations for more efficient workflows. Notable demonstrations featured companies like Supabase, Datadog, and Cloudflare, showcasing advancements such as natural language to SQL conversion and code generation, which proved to be more efficient than traditional tool calling. The Agentic AI Foundation, launching with 50 participating companies, promises to drive MCP's adoption as an industry standard, with its vendor-neutral governance removing barriers to enterprise adoption and fostering innovation in AI agent workflows.

On December 9, 2025, the Linux Foundation announced the creation of the Agentic Artificial Intelligence Foundation (AAIF), aimed at fostering open governance of three key projects: Anthropic's Model Context Protocol (MCP), Block's goose, and OpenAI's AGENTS.md. The AAIF seeks to ensure these projects remain open, community-driven, and vendor-agnostic, which is crucial as they become critical infrastructure for AI. With over 10,000 MCP servers now in operation across platforms such as Claude, Microsoft Copilot, and ChatGPT, the move to open governance mirrors past industry shifts toward shared standards, as seen with containers and Kubernetes. This initiative has garnered significant industry support, with AWS, Google, and Microsoft as Platinum members, highlighting a shift towards interoperability and enterprise confidence in adopting agentic AI. The AAIF is set to provide a collaborative platform for addressing security concerns and developing authentication, authorization, and audit trails, thereby advancing the maturity of the agent ecosystem. As open standards accelerate development, the transition from conversational AI to autonomous agents coordinating across systems is underway, supported by AAIF's transparent governance framework.

Cross App Access (XAA) is an OAuth 2.0 extension designed to address the shadow IT problem by enhancing enterprise governance over app-to-app connections, particularly in contexts involving AI applications. Traditionally, when an AI app connected to a tool like Asana or Zendesk, the OAuth handshake bypassed the enterprise Identity Provider (IdP), leading to compliance challenges and visibility gaps for IT administrators. XAA solves this by incorporating the IdP into the OAuth process, allowing centralized control over app permissions without additional user consent steps. This extension is based on the Identity Assertion JWT Authorization Grant (ID-JAG) specification, which enables the IdP to issue a short-lived identity assertion that verifies the app's permissions to act on behalf of a user at a resource app. By shifting the approval and governance responsibilities to the IdP, XAA provides enterprises with visibility into app-to-tool delegations, centralized management of approvals and revocations, and a more streamlined rollout process for AI integrations, aligning with modern enterprise security and procurement needs.

The Model Context Protocol (MCP) enhances client-server interactions, necessitating a client identity for token requests, traditionally managed by Dynamic Client Registration (DCR). DCR allows clients to self-register programmatically, enabling seamless connectivity in MCP's many-to-many ecosystem by eliminating pre-registration. However, DCR poses security risks like client impersonation and redirect URI vulnerabilities, leading to operational burdens as servers accumulate dynamic client records. To address these challenges, Client ID Metadata Documents (CIMD) were introduced, shifting client identification responsibility from servers to clients by using stable HTTPS URLs, reducing server-side state management, and providing better trust signals. CIMD has become the preferred method for MCP client registration due to its scalability and reduced maintenance, though DCR remains relevant for local clients lacking stable HTTPS origins or requiring enterprise policy control. Despite its initial prominence, DCR's limitations in a large-scale ecosystem necessitated the evolution to CIMD, which better aligns with MCP's dynamic and web-native client landscape.

The introduction of Client ID Metadata Documents (CIMD) as part of the OAuth protocol represents a significant shift in client registration, particularly for open ecosystems like the Model Context Protocol (MCP). CIMD replaces the traditional Dynamic Client Registration (DCR) model, which required every client to be pre-registered with an authorization server, with a stateless and scalable approach where a client is identified by a URL hosting its metadata in JSON format. This method allows clients to introduce themselves in any ecosystem without necessitating a permanent registration entry on each server, making it ideal for environments where a single AI client may need to connect to thousands of servers. The authorization server fetches and validates the metadata from the client's URL, ensuring security and identity control via domain ownership, thus mitigating risks such as client impersonation and SSRF attacks. By using a web-native identity mechanism, CIMD simplifies the client implementation process, maintains a single client identity across multiple servers, and eliminates the need for managing server-side credentials, making it a preferred choice in MCP's 2025-11-25 specification for dynamic client registration.

As the landscape of SaaS products evolves with the integration of AI copilots and autonomous agents, the importance of robust authorization systems like OAuth 2.0 and OpenID Connect (OIDC) has become paramount. These frameworks have transitioned from being infrastructure best practices to essential product capabilities, particularly for those aiming to scale and engage with enterprises or integrate AI. OAuth 2.0 focuses on authorization, ensuring controlled access to resources, while OIDC provides authentication, verifying user identities. Together, they offer a standardized method to securely delegate access and identity management, crucial for AI-driven environments where software acts on behalf of users. The emergence of AI necessitates granular and context-aware authorization, prompting a shift towards short-lived tokens and dynamic scope management to mitigate risks associated with powerful and long-lived OAuth tokens. In this new era, OAuth and OIDC are foundational for maintaining trust and security, enabling SaaS products to safely participate in the AI toolchain by ensuring verified identity, explicit consent, least-privilege access, and revocable authority.

The 2025 revision of the Model Context Protocol (MCP) introduces "Tasks," an experimental feature that transforms MCP from a synchronous model into a call-now, fetch-later protocol, allowing asynchronous operations. Tasks address common issues like timeouts and blocked sessions by standardizing long-running operations, thus enabling seamless interoperability among clients, servers, and SDKs. This upgrade allows for better parallelization, consistent progress reporting, and efficient handling of extensive workloads, such as ETL jobs and large file conversions, by introducing a cross-request asynchronous state machine. Although tasks are currently experimental, they are already being actively implemented across languages, as they are foundational for managing complex workflows and enhancing user experience. Security considerations are emphasized, as tasks are fetched later by task IDs, requiring robust authorization mechanisms to ensure safe and compliant execution. Overall, Tasks are set to become a fundamental aspect of MCP, offering a more powerful and flexible approach to managing asynchronous operations within the protocol.

Amazon Nova Forge, unveiled at AWS re:Invent 2025, is a groundbreaking enterprise AI capability that enables companies to create custom foundation models, called "Novellas," by integrating their proprietary data and domain knowledge directly into the model training process, rather than merely fine-tuning existing models. This approach, starting at approximately $100,000 per year, makes it feasible for mid-market companies to develop tailored AI solutions, previously a realm reserved for frontier labs with extensive budgets. Nova Forge offers an open training framework with access to pre-training, mid-training, and post-training checkpoints, allowing enterprises to maintain their proprietary data within AWS boundaries while preventing catastrophic forgetting. Reddit's successful consolidation of multiple machine learning models into a single, more efficient Nova-based model exemplifies Forge's potential to streamline operations and enhance performance. The platform's implications extend across various industries, including life sciences, finance, and manufacturing, where deep domain knowledge and proprietary data are crucial. By lowering the barrier to custom model development, Nova Forge democratizes AI capabilities, prompting enterprises to evaluate whether their unique data justifies building bespoke models that could surpass generic alternatives.

WorkOS Connect has introduced Client ID Metadata Document (CIMD) support for MCP authentication, allowing clients to identify themselves to authorization servers securely without pre-registration by providing an HTTPS URL as a client_id. This new feature complements the existing dynamic client registration (DCR) method. Additionally, a consent screen has been added to Single Sign-On (SSO) for enhanced security against CSRF and phishing attacks, prompting users to confirm their profile information. The release of the AuthKit SDK for TanStack Start offers a comprehensive suite of tools for implementing secure, server-first authentication in Start apps, including middleware and utilities for handling sessions and redirects. Furthermore, WorkOS now supports SSO connections through the Clever identity platform using OIDC, with simplified setup instructions available in the Admin Portal. Other updates include collaboration with Microsoft on the Entra Agent ID development and an updated Developer’s Guide to SAML authentication, providing detailed guidance on various aspects of SAML.

MCP's 2025 release introduces a significant change in the OAuth client registration process, shifting from Dynamic Client Registration (DCR) to Client ID Metadata Documents (CIMD). This transition aims to address the scalability issues inherent in DCR, where clients register with multiple servers, leading to potential security risks and administrative overhead. CIMD offers a solution by allowing clients to identify themselves through a URL pointing to a JSON document, streamlining the process by eliminating the need for per-server registration and reducing client sprawl. This approach aligns with MCP's scale and decentralization needs, as it leverages web-native identity patterns, enhances security by preventing unauthorized access, and simplifies client management without relying on manual registration. While DCR remains useful in controlled environments requiring strict client oversight, CIMD is now the recommended default for MCP, with WorkOS already supporting this new method to facilitate smoother OAuth integration.

Duality AI's Falcon platform simulates high-fidelity real-world environments, allowing teams to develop and test AI and robotic systems in virtual "digital twins" without the risks or costs of physical trials. This approach accelerates model training and testing by generating synthetic perception data and enabling autonomous behavior testing, with recent enhancements like FalconEditor and FalconCloud broadening accessibility and usability. While Falcon provides the simulation and data generation capabilities, WorkOS complements it by offering enterprise-grade identity and access management solutions, including user authentication, authorization, and governance, crucial for securely managing and deploying the resulting simulations and data across various teams and clients. This integration forms a comprehensive stack for developers and product builders, facilitating rapid AI and robotics development while ensuring secure access and compliance, making it particularly valuable for industries dealing with complex automation and AI-driven solutions.

OAuth 2.0 is an authorization framework that enables applications to access a user's data on another service without needing the user's password, using short-lived tokens with specific permissions. It is widely used in scenarios like third-party integrations, account connection flows, and machine-to-machine access while ensuring security through scoped, time-limited tokens. OAuth 2.0 is not meant for authentication, which requires OpenID Connect. Key flows include Authorization Code with PKCE for user logins, Client Credentials for service-to-service access, Device Code for constrained devices, and Refresh Token for maintaining long-term access. Despite its utility, OAuth 2.0 can be complex to implement correctly, with pitfalls such as insecure token storage and over-broad scopes. OAuth 2.1, still a draft, incorporates modern best practices and simplifies implementation. Providers like WorkOS offer managed solutions to streamline OAuth integration by handling complexities and ensuring secure, scalable implementations, thus allowing developers to focus on product development rather than protocol intricacies.

Authorization in modern software is a complex, rapidly evolving challenge, often starting with basic Role-Based Access Control (RBAC) but quickly outgrowing its simplicity as Software as a Service (SaaS) products scale and evolve. Traditional models like RBAC and Fine-Grained Authorization (FGA) often struggle to keep pace with the dynamic nature of SaaS products, leading to frequent rewrites and increased complexity. WorkOS offers a flexible authorization system that extends the basic RBAC model, introducing a fine-grained approach that naturally evolves with application structures without requiring new languages, schema migrations, or major rewrites. This approach supports nested, resource-scoped permissions, allowing for automatic inheritance and scoping across hierarchical resources, thereby aligning authorization models with real product architectures. WorkOS’s solution integrates seamlessly with existing identity systems and supports enterprise requirements, including automated access control and AI-driven features, ensuring fast, consistent, and scalable performance while reducing the need for bespoke authorization systems.