MCP Weekly: Anthropic's Constitution, GitHub Copilot SDK, and OpenAI's $30B Bet

Welcome to the latest installment of the MCP Weekly, covering major developments from January 21st to January 28th. As AI systems move from chat to action, this week brought clarity on governance, developer access, and the capital flowing into the space.

TL;DR

Anthropic released an 84-page Constitution establishing priority rules for Claude's behavior and donated it to the public domain. GitHub launched the Copilot SDK in technical preview, allowing developers to embed the agent core into any application. SoftBank is in talks to invest $30 billion in OpenAI at an $830 billion valuation, following its $22.5 billion investment last month.

Major Updates of the Week

Anthropic Constitution for Claude and Public Sector AI Deployment

Anthropic has reinforced its commitment to responsible AI and high-stakes deployment with three major announcements:

• Claude's Constitution: An 84-page, CC0-licensed Constitution was published for its Claude models, establishing a four-tier priority hierarchy (safety, ethics, compliance, helpfulness). The document shifts from rigid rules to a "virtue-based" judgment approach, allowing Claude to handle complex scenarios while maintaining hard constraints against high-risk behaviors.
• UK Government Partnership: Anthropic was selected by the UK Department for Science, Innovation and Technology (DSIT) to pilot an "agentic" Claude-powered assistant for the GOV.UK portal, with an initial focus on personalized employment advice. The initiative includes knowledge transfer to build in-house UK technical expertise and prioritizes user privacy with full data control and opt-out options.
• Claude’s Productivity Tool Integration: Claude was transformed into a centralized work hub by integrating nine productivity tools (including Asana, Slack, and Figma) via the MCP standard. This allows users on Pro, Team, and Enterprise tiers to perform real-time, bi-directional actions (like drafting in Clay or creating diagrams in FigJam) through live components without leaving the chat interface.

‍GitHub Copilot SDK Launches in Technical Preview

GitHub introduced the technical preview of the GitHub Copilot SDK, allowing developers to embed Copilot's agent core into any application. Supporting Node.js, Python, Go, and .NET, the SDK enables multi-step planning, custom tool invocation via the MCP, and file system commands. It also supports integration with OpenAI's GPT-5/GPT-5.2 and Anthropic's Claude 4 series.

OpenAI Technology and Infrastructure Updates

OpenAI made three key announcements detailing new products and engineering strategies:

• Prism for Scientific Research: Introduced Prism, a free, LaTeX-native workspace powered by GPT-5.2 that consolidates drafting, revision, and collaboration for researchers. It is built on the acquired Crixet platform and features deep context awareness, native LaTeX workflow support, and the ability to convert handwritten equations into formatted LaTeX.
• Codex Agent Loop Deep Dive: Published a technical breakdown of the "agent loop" powering its Codex CLI (a local software agent for autonomous code editing). This harness orchestrates a continuous cycle of user input, GPT-5.2 inference, and tool execution, featuring dynamic prompt construction, prompt caching for performance, and automatic compaction to summarize long threads.
• PostgreSQL Scaling Strategy: Revealed its technical approach for scaling PostgreSQL to support 800 million ChatGPT users. The strategy uses a single-primary architecture with nearly 50 globally distributed read replicas, offloads write-heavy workloads to Azure Cosmos DB, and employs PgBouncer and cache locking to maintain low-latency, high-availability connections.

SoftBank in Talks for $30 Billion OpenAI Investment

SoftBank is in advanced talks for a $30 billion investment in OpenAI, as part of a potential $100 billion funding round valuing the company at $830 billion. This follows SoftBank's $22.5 billion investment in December 2025 for an 11% stake. To finance this, SoftBank sold its $6 billion Nvidia stake. The new funding is linked to the $500 billion Stargate initiative for AI data centers and renewable energy.

Ai2 Releases Open Coding Agents and SERA Framework

Ai2 introduced Open Coding Agents and the SERA framework for specialized coding agents on private codebases, reducing development costs by up to 57x. The SERA-32B model, which uses partially correct patches to avoid expensive testing, achieves a 54.2% solve rate on the SWE-bench Verified benchmark, outperforming prior open-source models. Reproducing performance costs about $400, and repository specialization around $1,300. Ai2 released the models, synthetic trajectories, training recipes, and a Claude Code-compatible CLI.

Other Updates

Google announced on Jan 24th 2026 the integration of Datadog and Google’s Agent Development Kit (ADK) to provide automatic instrumentation for systems built with ADK and it gives the visibility to monitor agent behavior, to track costs and errors, and optimize agents for response quality.

pgEdge announced significant updates to its Postgres MCP Server, part of the pgEdge Agentic AI Toolkit. This server allows AI assistants to interact with Postgres databases using natural language. Beta 2 focuses on moving beyond simple queries to allowing AI to actively manage and modify data. Beta 3 introduced more features for developers.

My Thoughts: Governed Action Is Now the Baseline

AI is maturing in governance, as shown by Anthropic's new public 'constitution,' which sets a baseline standard for other companies. GitHub's open-sourcing of the Copilot SDK democratizes agent logic, shifting focus from orchestration to domain expertise. SoftBank's investment in full-stack AI development, from models to data centers, highlights the need for reliable, scalable, and trustworthy systems operating under governance. Standards like MCP are key to this transition.

‍