Claude Execution Lab
A visual simulator for three common agentic workflow patterns — Claude Code, MCP tool calls, and Agent Skills invocations. Select a mode, run the simulation, step through each phase, and inspect what each step does and why it matters. Educational visualization by Eggthropic. Does not reveal Claude's private reasoning or internal architecture.
A simulation of how Claude Code might decompose a refactoring task — reading files, planning a split, editing code, and running checks.
Not Claude's chain of thought, an internal debug trace, or a guarantee that real executions follow these exact steps.
Refactor Hero.tsx into HeroHeadline and HeroCTA components
Refactor a landing page component
Press Run Simulation to start the trace.
Run the simulation to see step-by-step context here.
Claude Code requires explicit permission grants for file writes. Defaults are conservative — review before expanding.
Shell commands have broad blast radius. Confirm before granting unrestricted bash access to any agent task.
Always review the full diff before committing. Generated code can be plausible but logically incorrect.
Local refactors operate on local files only. No external network calls required in this scenario.
File edits are reversible with git. Commit or stash before starting agent tasks so you can roll back.
Levels are illustrative — actual risk depends on your specific environment, permissions, and threat model.
Agentic systems make decisions and take actions — but human review before committing, merging, or shipping remains essential.
See Trust & Safety →MCP is an open protocol. Every MCP server is external code that your agent trusts. Audit sources, limit permissions, validate responses.
MCP Visual Explainer →A skill packages a workflow into a directory with a SKILL.md. It's not a plugin marketplace — it's structured instructions the agent loads on demand.
Agent Skills Explainer →Lint and build passing means the code compiles. It does not mean it is correct, secure, or matches your intent. Always read generated diffs.
Claude Code experiment →What is the Claude Execution Lab?
The Claude Execution Lab is an interactive educational visualization built by Eggthropic. It demonstrates how agentic AI workflows are commonly structured — covering three distinct patterns: Claude Code for code editing tasks, MCP (Model Context Protocol) for external tool calls, and Agent Skills for invoking reusable task workflows. All simulations run locally in your browser using predetermined trace data. No real Claude API calls are made.
The goal is to make agentic workflows legible — to show the phases, the decisions, the risks, and the review points that matter — before you build with these systems yourself.
Claude Code workflow simulation
The Claude Code simulation shows a typical agentic coding task: refactoring a landing page component. The simulated workflow reads relevant files, plans the decomposition, loads edit tools, writes the changes, runs lint and build checks, and produces a diff summary.
Key phases
- Read files — build context before making any changes
- Plan refactor — separate reasoning from execution
- Edit code — create new components and update imports
- Lint & build — automated checks for obvious regressions
- Summarize — diff report for human review
Claude Code requires explicit permission grants before reading or modifying files. Defaults are conservative. Always review generated diffs before committing.
MCP workflow simulation
The MCP simulation shows how a Claude-based agent might use the Model Context Protocol to fetch live product data from an external tool server. The workflow parses a natural language request, connects to an MCP server over Streamable HTTP, fetches the tool schema, calls the tool, and validates the response.
About MCP
MCP is an open protocol — not Anthropic-exclusive — based on JSON-RPC 2.0. It defines three primitives: Tools (functions the model can call), Resources (structured data), and Prompts (reusable templates). The official MCP Registry at registry.modelcontextprotocol.io launched in September 2025. The current remote transport is Streamable HTTP; SSE transport was deprecated in the 2025-03-26 spec revision.
Security note: treat every MCP server as untrusted unless you control or have audited the source. Malicious servers can return prompt-injection payloads or exfiltrate sensitive context.
Agent Skills workflow simulation
The Agent Skills simulation shows how Claude Code might invoke the /pr-describe skill — scanning the skills directory, loading the SKILL.md instructions, running an allowed helper script against the git diff, and generating a structured PR description.
About Agent Skills
An Agent Skill is a directory containing a SKILL.md file with YAML frontmatter specifying the skill name, description, and allowed tools. Skills are invocable as slash commands in Claude Code and are supported in the Claude Agent SDK. They are loaded from .claude/skills/ in the project workspace or installed from the official anthropics/skills repository. Behavior may vary by Claude surface and configuration.
Why human review still matters
Agentic workflows compress time and reduce mechanical work — but they do not eliminate the need for human judgment. Lint passing does not mean code is correct. Validated MCP responses can still contain wrong data. Structurally complete skill outputs can still miss context only you have.
The Trust & Safety panel in this lab surfaces the review points for each workflow mode. Use them as a checklist, not as reassurance that the work is done.
Related Eggthropic labs
Frequently asked questions
Is this connected to the real Claude API?
No. The Claude Execution Lab uses simulated execution traces only. No API calls are made, no keys are required, and no real model inference runs in the browser.
Does this show Claude's private reasoning?
No. It visualizes external workflow patterns that are common in agentic systems — not private chain-of-thought, internal model state, or any form of actual Claude reasoning.
Can Claude Code actually edit files and run commands?
Claude Code can work with codebases and perform actions such as reading files, making edits, and running shell commands — depending on your environment configuration, permissions granted, and explicit user approval at each step. Real behavior depends on your setup.
Is MCP only for Anthropic?
No. MCP (Model Context Protocol) is an open protocol for connecting AI systems to external tools and data sources. It is not Anthropic-exclusive — it has broad adoption across the AI ecosystem and official SDKs for all major languages.
Are Agent Skills the same as prompts?
No. Agent Skills package task-specific instructions, supporting scripts, and resources into reusable directory-based capabilities invocable as slash commands. They are more structured than a prompt and can include helper scripts. Availability and behavior can vary by Claude surface and setup.