Anthropic adds code execution, MCP connector, Files API, and caching

Anthropic announced four API capabilities for agents: a code-execution tool, an MCP connector, a Files API, and extended prompt caching.

By Hypler editorial team · Source: Anthropic.

Original source date: .

Hypler publication: . Hypler updated: .

Estimated reading time: 4 minutes.

Topics: MCP, code execution, files

Anthropic logo

Generated illustration of modular compute gates and an isolated coding workflow.
Generated illustration of modular compute gates and an isolated coding workflow. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

Four launches, four control problems

Anthropic presented code execution, an MCP connector, a Files API, and extended prompt caching together on May 22, 2025. The grouping makes product development look coherent: computation can run away from the application process, tools can be reached remotely, files can persist, and long context can be reused. Operationally, those are different trust boundaries. The release did not create a universal security envelope, nor did it establish current entitlements, regional support, retention, billing, or safe composition for every account and workflow.

One agent can cross several planes

A practical design separates reasoning, sandbox, connector, and evidence planes even when one model session coordinates them. Code execution needs filesystem and egress limits. An MCP connection needs a verified server identity, a scoped manifest, schema checks, and a revocation path. A file needs an owner, classification, lifetime, and deletion receipt. A cache key needs tenant, authorization, and content-version dimensions. Text generated by a model may request work across those planes; it cannot grant the permissions that make the request possible.

Composition is where the exposure grows

Each component can appear bounded alone and still create a wider pathway in combination. A hostile filename can influence a model, a sandbox can export a connector response, and a stale cache can surface context after access changed. Ingestion must treat file content as untrusted. Sandboxes should lack unapproved network paths. Connector calls should not automatically consume arbitrary outputs from code. A partial failure also needs an honest terminal state. An operator should be able to tell whether an answer came from a file, a cached prefix, computation, a tool response, or a failed fallback.

A repair request keeps the seams visible

Hypothetical example: imagine an engineer asking for candidate repairs to malformed records. A bounded export enters a file service with a classification label. A no-network sandbox calculates possible fixes, while the model explains the patterns. The only connector available creates a review ticket; it cannot update the source system. That ticket names the file version, ruleset version, sampled failures, and a patch preview. A separate service-owned apply job waits for review. Missing classification stops the run early, and a connector error reports that no ticket was created.

The test suite needs cross-plane cases

Adoption evidence should include expired grants, revoked connector credentials, mislabeled or oversized files, cross-tenant cache attempts, malicious filenames, sandbox egress, connector schema drift, and deletion after a partial failure. A document that asks the model to call an unrelated MCP tool is especially useful: the file, sandbox, and tool policy should all contain it. Tests should establish that unavailable tools cannot be discovered or invoked and that cache entries inherit authorization and content revisions. Happy-path text output is a weak substitute for those containment proofs.

Inventory turns components into accountable systems

A machine-readable inventory can list every sandbox image, connector, file store, cache namespace, and model-facing grant with its owner, data classes, authentication method, egress policy, expiry, logging location, and test evidence. Review it when an integration or version changes. The operational questions then become answerable: what could this run touch, what did it touch, what was denied, and how is access revoked? That is the discipline the vendor grouping leaves to implementing teams.

Hypler’s judgment on the bundle

The useful takeaway is not that agents need more standing authority. It is that separate grants make powerful combinations inspectable. Hypler evaluates files, computation, connectors, and evidence as distinct capabilities that can be observed and reviewed within one workflow. That keeps a productive integration from quietly becoming a broad privilege bridge.

Hypler analysis

Hypler’s view

Combining these primitives can concentrate powerful permissions. Sandboxes, connector identity, file lifecycle, data retention, and cache invalidation need separate controls before an agent can safely coordinate them.

Hypler technical analysis

Engineering notes

hosted code execution, remote connectors, persistent files, and context caching are separate capability planes that compound risk when joined by one agent. Each plane needs its own grant and evidence: sandbox filesystem and egress limits; connector identity, manifest, schema, and revocation; file ownership, classification, lifecycle, and deletion receipt; cache key dimensions for tenant, authorization, and content revision. A model session may coordinate work, but it cannot grant itself access across planes. The output trace should identify whether information came from an uploaded file, code result, connector, or cache, and partial failures must not imply success. A machine-readable inventory lets operators see what a run could touch, did touch, and was denied.

What to inspect

Capability-grant inventory
Inspect per-run sandbox limits, connector identity and manifest, file classification and lifecycle, cache namespace, authorization dimensions, owner, and revocation path.
Cross-plane adversary
Use a hostile filename or document that requests an unrelated connector call or network export; inspect containment at file ingestion, sandbox, and tool policy.
Lifecycle failure drill
Test expired grant, revoked connector, mislabeled or oversized file, cross-tenant cache attempt, and deletion after partial failure; inspect honest terminal states.

Scope and limits. The grouped release does not create one universal security boundary or establish account entitlements, retention, billing, regional access, or safe composition by default.

Terms and definitions

MCP
Model Context Protocol, an open protocol for connecting AI applications to tools, data sources, and prompts through defined interfaces.
Tool Calling
A model interface pattern where the model requests a structured operation from an external tool.
API Layer
The server-side interface that exposes defined data and actions to clients or integrations.
Testing and Observability
Testing checks expected behavior; observability uses runtime signals to understand a system’s actual behavior and condition.

More on this subject

  • Anthropic adds remote MCP support to Claude Code

    Anthropic announced remote Model Context Protocol support in Claude Code. The company said developers could connect Claude Code to remote MCP servers by adding a vendor URL instead of operating local servers.

  • Anthropic introduces prompt caching

    Anthropic introduced prompt caching for the Claude API, describing reuse of stable prompt prefixes to reduce cost and latency on repeated long-context requests.

  • Anthropic introduces web search on the API

    Anthropic introduced web search as an API tool for supported Claude models, returning web results and citations for applications using current public sources.

Generated illustration of isolated tool-execution fixtures and a controlled test gate; not deployed software.
Generated illustration of isolated tool-execution fixtures and a controlled test gate; not deployed software. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

Original source

Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.

Related capabilities