News

Anthropic published a guide for wiring Claude Managed Agents into Vercel's Chat SDK. Each side owns half: Managed Agents runs the model, the tool loop, session state and sandboxed web research on Anthropic's servers; Chat SDK is the chat layer — one type-safe handler behind 15+ platform adapters (Slack, Teams, Discord, WhatsApp, web).

Together you get a research bot that streams its answer token by token, with a live feed of the tool calls it is making — and no database of your own, because the conversation lives in the Managed Agents session.

It needs Node 22.9+ and an Anthropic API key; `npm run setup` mints the agent and environment ids. One catch the docs are explicit about: it holds a long-lived connection for a whole turn, so it does not fit serverless platforms with second-level timeouts.

Billing is per API usage, separate from any subscription. It earns its keep when real people keep asking the same class of question in a chat tool — before that, it is a toy.

A new `--restricted` flag (or `CLAUDE_CODE_RESTRICTED=1`) strips out every built-in tool that runs commands or code, drops WebFetch, confines file tools to the working directory, refuses `bypassPermissions`, and ignores your user, project, and local settings files.

It answers a real worry: you want to read unfamiliar code with an AI, but you don't want that AI executing anything on your machine. Until now you had to catch each attempt at the permission prompt. Now you can remove the capability outright.

Run `claude --restricted`, and name any tool you do want back with `--tools`.

Cross-session messaging (`SendMessage` / `ListAgents`) now works on Bedrock, Vertex, and Foundry, and in setups with telemetry disabled — none of which were covered before.

What it's for: run two terminals, one planning and reviewing, one writing code, and let them talk directly instead of relaying through you. The same release collapses incoming peer messages to a one-line preview (Ctrl+O expands) so a long message doesn't flood your screen.

This site was built that way — one window named plan coordinating, another doing the edits.

`/permissions` gained an Auto mode tab where you can read and edit the classifier rules — until now, what auto mode let through was invisible. The same batch added a one-keystroke "Yes, and switch to auto mode" option on Bash permission prompts.

Worth noting alongside it: Bash allow rules with a wildcard before the subcommand (e.g. `Bash(git * main)`) now warn at startup, because they also match options inserted before the subcommand — wider than you probably meant.

If permission prompts interrupt you constantly, read the Auto mode tab first, then switch once you're comfortable with what it covers.

A new `/claude-api cost-optimize` profiles an existing project's Claude API spend and walks you through the levers one at a time: caching, token hygiene, batching, effort, model choice — one measured change per step.

The value isn't the phrase "save money." It's the enforced order. Most people change five things at once and never learn which one mattered.

Only relevant if you're building on the Anthropic API yourself; subscription-only Claude Code users can skip it.