Anthropic adds code execution, MCP connectors, and 1-hour caching to its API

Building a real AI agent used to mean writing miles of unglamorous glue code. You had to spin up secure Python sandboxes, maintain client harnesses for external integrations, and juggle file uploads on every single turn.

Anthropic just rolled out four major upgrades to its API that handle all that heavy lifting natively. The update includes a sandboxed code execution tool, an MCP connector, a dedicated Files API, and extended prompt caching.

Combined with Claude Opus 4 and Sonnet 4, these beta features essentially turn Claude from a raw text generator into a full-fledged execution engine right out of the box.

Native Python and zero-friction integrations

Instead of Claude just spitting out a Python script for you to run locally, the new code execution tool runs Python code directly inside an isolated sandbox.

Claude can now iteratively clean datasets, run financial modeling, execute statistical tests, and generate charts in a single request. If code fails, it can refine and rerun it until it works. Anthropic gives teams 50 free container hours per day, after which it costs $0.05 per container hour.

Alongside code execution, the new Files API lets developers upload documents once and reference them across multiple sessions. Instead of repeatedly paying tokens to send a giant dataset on every prompt, you upload it once, and the code sandbox reads and writes directly to those files.

Meanwhile, the MCP connector removes the friction of building custom client infrastructure for the Model Context Protocol. Developers simply pass a remote MCP server URL, like Asana or Zapier, directly in their API call. Claude handles tool discovery, execution reasoning, and authentication under the hood.

To keep long-running agents practical, Anthropic also introduced extended prompt caching. While standard cache expires after 5 minutes, teams can now opt into a 1-hour TTL. For multi-step workflows, Anthropic claims this can cut long-prompt costs by up to 90% and slash latency by up to 85%.

Why it matters

We think Anthropic is systematically turning Claude from a conversation partner into an autonomous cloud worker.

For builders, this drastically lowers the barrier to deploying real agents. You no longer need third-party sandbox services or complex agent frameworks just to let Claude crunch a spreadsheet and update a project board. With a daily 50-hour free tier for code execution and an hour of cached memory, running continuous analytical tasks is now dirt cheap.

For the broader AI ecosystem, this piles immense pressure onto orchestration wrappers and sandbox startups. OpenAI had a head start with its Code Interpreter and Assistants API, but Anthropic is coupling native execution with its rapidly growing, open MCP standard.

It is a very smart play for developer lock-in. But it also sends an unmistakable signal to the tooling layer: if your product merely manages prompts, tools, or sandboxes, foundation model APIs are coming for your lunch.