Claude 4 - First Tests and Impressions (Claude Code, MCP API, Code Execution Tool++)

Claude 4’s new tooling is getting a hands-on stress test: Anthropic’s Claude Code can now run Python inside a secure sandbox via a “code execution tool,” and the Claude API can connect to external capabilities through MCP servers. The practical takeaway from these first experiments is that the workflow is fast to wire up, produces usable execution reports, and can be chained into larger apps—while cost remains relatively modest for small runs.

The testing begins with Claude Code identifying which Claude 4 model is active, then moves into a sandbox execution trial. A Cloud Sonnet agent brainstorms a solution to a user problem, and the proposed Python code is sent to Claude Opus for execution in the sandbox. The example generates 20 random numbers, computes mean, median, and standard deviation, builds a histogram, and flags outliers. The execution tool returns a structured “final report” with workflow success, analysis, edge cases and limitations (like empty lists and single elements), and a code-quality assessment. The experimenter then runs the generated code locally multiple times to verify behavior: the mean varies run-to-run but trends toward the expected value when scaling up to 1,000 random samples (hovering near 50). A quick cost check via “/cost” lands at roughly “almost $1,” suggesting the sandbox execution is not prohibitively expensive for iterative development.

Next comes a more ambitious app built with Claude Code and Cloud Sonnet: a webcam-based object-to-poem-and-voice pipeline. The system captures objects from a live camera feed, identifies them, and uses Claude Opus 4 to write a short poem once three objects are detected. That poem is then sent to the ElevenLabs API for text-to-speech using a configured voice ID, and the audio is played back in the UI. Early runs show the concept working quickly, but object identification based on a weaker approach mislabels items (e.g., confusing a sports ball). The app is then upgraded to use Claude’s vision capabilities for object identification, with the workflow adjusted to take a screenshot, extract object names from the vision response, and populate the list. Subsequent tests improve accuracy: a protein bar is recognized, Bose headphones are identified, and a coffee mug is correctly labeled; the resulting spoken poems match the detected objects and arrive rapidly, implying asynchronous handling in the app.

Finally, the MCP connector is demonstrated as a plug-in style integration. Using an MCP server URL (sourced from mcpservers.org), Claude Opus is prompted to compute a math expression via a “sequential thinking” MCP tool, producing a step-by-step order-of-operations breakdown (PEMDAS) and a verified result. The experiment then swaps the MCP server to a “fetch” tool, showing that changing the MCP server in the API configuration can redirect Claude from reasoning tools to content-fetching tools with minimal code changes.

Overall, these first impressions frame Claude 4’s value less as a single model upgrade and more as an orchestration layer: sandboxed execution, vision-driven app logic, and MCP-based tool connectivity combine into workflows that can iterate quickly, verify outputs, and expand into external services.