Anthropic takes the toolchain —
Stainless, KPMG, and 20+ legal MCPs.

Stainless turns an OpenAPI spec into idiomatic TypeScript, Python, Go, Java and Kotlin SDKs — plus CLIs and MCP servers — and has powered every official Anthropic SDK from day one. It also powers OpenAI's, Google's and Cloudflare's. The Information puts the price north of $300M. Hosted Stainless products wind down; customers keep the SDKs they've already generated. Strategically this is the most consequential move of the quarter: every other lab now either rebuilds, migrates, or ships behind Anthropic's release cadence on the most boring-but-essential piece of the agent stack.

A week after the PwC alliance expanded to 30,000 trained professionals, KPMG signed a global alliance to embed Claude Cowork and Managed Agents into Digital Gateway, its Microsoft Azure-hosted client platform, and roll Claude to its entire 276,000-person workforce. Tax & Legal go first, with full Azure-side deployment targeted for September. The strategic read: Anthropic is locking the Big Four one deal at a time, and the contract size is now measured in workforces, not seats. Whatever Claude's next model number is, the distribution moat is what's being built right now.

Anthropic wired Claude into roughly every part of the legal stack at once: Ironclad, iManage, NetDocuments, DocuSign, Relativity, Everlaw, Consilio, Box, Datasite, Harvey, Thomson Reuters CoCounsel, plus newer research engines (Midpage, Trellis, Legal Data Hunter). On top sits twelve practice-area plugins — Corporate, Litigation, Privacy, IP, Employment, Regulatory, AI Governance and others — each opened with a setup interview that learns the team's playbooks, escalation chains and house style. Pair this with the small-business workflows and KPMG: Anthropic's go-to-market shape is now "vertical Claude × deep distribution partner", not horizontal platform.

Two features that quietly unblock the enterprise version of "agents that touch real systems". MCP tunnels (research preview) route Managed Agents to MCP servers inside private networks without exposing them publicly. Self-hosted sandboxes (public beta) move tool execution into the customer's own infrastructure — or a partner's (Cloudflare, Daytona, Modal, Vercel) — while Anthropic still runs the orchestration loop. The shape is the same as the Antigravity-managed pattern Google announced last week; the difference is that Anthropic is giving you the seam to put the execution wherever your security team can sign off on it.

Anthropic reversed an earlier gag on its Mythos cybersecurity preview: partners in Project Glasswing (Amazon, Microsoft, Nvidia, Apple and others) may now disclose their participation and share findings, tools and code with peer security teams, regulators, open-source maintainers and the press, under normal responsible-disclosure norms. Palo Alto Networks and Mozilla had already credited Mythos with surfacing volumes of vulnerabilities they couldn't get to via standard processes; this change is what turns those findings into actual industry signal instead of NDA shelf-ware.

Released the same day as the Antigravity 2.0 launch, ADK 2.0 promotes Google's Agent Development Kit from "framework you trial" to "framework you ship on". The headline addition is a model-agnostic multi-agent workflow engine — non-linear, conditional and cyclical execution patterns, native agent-to-agent routing, nested hierarchies, and parallel sub-agent workers. A day earlier, v1.34.0 added persistent task stores, Gemini Live API support in evaluate, mTLS Cloud Telemetry and a Skill Registry. ADK is what Google now plugs into Antigravity, Gemini Enterprise and the Managed Agents API — not three separate runtimes.

The headline feature is the unsexy-but-load-bearing one: error handlers that resume across host crashes via the durable execution path, not just retries inside a single process. 1.2.0 also adds set_node_defaults() for less boilerplate, delta channel snapshots, and a coordinated bump of langgraph-prebuilt, langgraph-cli, and the SQLite/Postgres checkpoint packages out of alpha. If your agent is something you need to babysit through a deploy, this is the release to read.

Five releases in seven days. The one to note is v2.1.142 (May 14): Opus 4.7 replaces 4.6 as the default model for Fast mode, with a CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE pin for teams that want to hold. The rest is the usual cadence — claude agents --json for scripting live sessions, /resume for background sessions, plugin dependency enforcement, projected context cost in the plugin marketplace, worktree.bgIsolation, PowerShell -ExecutionPolicy Bypass on Windows, status-line JSON exposing repo/PR. Claude Code is now shipping like a daily product, not a release-train tool.

Foundry Hosted Agents samples for RAG, Skills and Memory; better intermediate-output handling for workflows and orchestrations; YAML block-scalar parsing in SKILL.md frontmatter; tools added by ContextProvider.before_run now make it into session creation; and agent-framework-orchestrations moves to RC. A quiet but pointed release: Microsoft is converging the framework on the same primitives — skills, memory, hosted agents — that Anthropic and Google are shipping their platforms around.

SQLite at ~/.dejavu/, no cloud, no account, four surfaces (Python, CLI, REST, MCP) so the same memory file is reachable from Claude Desktop, Cursor and anything else that speaks MCP. The pitch is that "agent memory" doesn't need to be a managed service — for a single developer, a local file plus an MCP server is enough, and avoids the data-residency problem most teams quietly have with hosted memory.

A local Rust daemon that replaces cat and grep for coding agents with hybrid BM25F + semantic search over AST chunks, session-level dedup, and an MCP front end. Reported read-token savings of 84–88% on the kind of "read these 30 files" task agents waste their context on. Same throughline as Semble and Needle the last two editions — the small-tools-with-big-token-savings genre is the most consistent OSS pattern of the quarter.

A Rust + TypeScript desktop app that reads your local skill files, fuzzy-searches across names and descriptions, groups by category, surfaces usage stats from session history, and one-clicks the matching slash command into Claude Code. Offline, MIT-licensed, no telemetry. The release matters more than the feature list: as skills proliferate (and they are — see the Watch list) the missing piece is a way to actually navigate your own library, and the first credible one has now shipped.

Apache-2.0 TypeScript framework that runs 67 named specialists through evidence-backed debate rounds, with mandatory human gates and a 10-pass verify loop. Supports Anthropic, Mistral (for EU deployments) and local Ollama; ships a "Clawern" daemon mode for continuous multi-client document processing. The author is unusually honest that the output-quality claim is a hypothesis, not a benchmark. Worth tracking against Anthropic's official legal stack above — same problem, opposite philosophy.

A CBOR-encoded protocol explicitly framed as complementary to MCP: sub-50-byte frames, under 16KB MCU footprint, validated on ESP32-class hardware. A reference Bridge translates DCP ↔ MCP, so any MCP-compatible LLM can drive a physical device without modification. Tiny project, but it's the first serious attempt to extend the agent-tool protocol family down the cost curve — same way HTTP quietly stopped being a desktop-only protocol once it ran on routers.

ClipboardHealth open-sourced the CLI they use in-house: pick tickets out of Linear, spawn one git worktree per ticket, hand each to Claude Code or Codex, sandbox by default (Safehouse on macOS, Docker on Linux). The "fleet of agents working off the backlog" pattern has been promised for a year by the managed-agent vendors; this is the same idea you can run on your own laptop tonight, against the project tracker you already use.