Anthropic Administrator

@Anthropic Administrator list all available models and show me which ones support vision capabilities

Based on the current model catalog, here are the available models with vision support: Claude 3.5 Sonnet (claude-3-5-sonnet-20241022) supports vision with a 200K context window, Claude 3 Opus (claude-3-opus-20240229) supports vision with a 200K context window, and Claude 3 Haiku (claude-3-haiku-20240307) supports vision with a 200K context window. All vision-capable models can process images alongside text in the same request. Non-vision models like Claude 2.1 remain available for text-only workflows.

This example demonstrates model discovery, which is essential before making API calls. The MCP returns live data from Anthropic's model catalog, so availability and specifications may change as new models are released or deprecated. No API key permissions beyond basic read access are required for listing models.

@Anthropic Administrator create a message using Claude 3.5 Sonnet to summarize the key risks in our Q4 product roadmap document

Message created successfully using claude-3-5-sonnet-20241022. The model identified three primary risks: (1) dependency on third-party API stability for the new integration feature, (2) aggressive timeline for the mobile redesign given current team capacity, and (3) unclear success metrics for the experimental AI features. The response consumed 1,247 input tokens and generated 312 output tokens. Total cost for this completion was approximately $0.015.

This example shows the MCP acting as a proxy to Anthropic's Messages API, allowing you to invoke Claude models directly from a Space. Each invocation consumes tokens from your Anthropic account and incurs standard API costs. The MCP requires a valid API key with message creation permissions. Token counts and costs are returned for budget tracking.

@Anthropic Administrator set up prompt caching for our standard system prompt about code review guidelines, then use it to review this pull request

Prompt cache created successfully. Your 2,400-token system prompt (code review guidelines, security checklist, and style preferences) is now cached and will be reused across subsequent requests for the next 5 minutes. Applied the cached prompt to analyze the pull request: the code follows naming conventions and includes appropriate error handling, but the database query in line 47 should use parameterized statements to prevent SQL injection. Estimated cost savings: 90% reduction on input tokens for the cached portion across repeated reviews.

This example highlights prompt caching, which significantly reduces costs and latency when you have static prompt content reused across multiple requests. Caches expire after 5 minutes of inactivity, so this works best for batch operations or active review sessions. The MCP handles cache key management automatically, but you're responsible for structuring prompts to maximize cacheable content.

When you need programmatic Claude calls inside your workspace

A 6-person engineering team building an internal code-review bot wants to call Claude models directly from their Switchy workspace without leaving the conversation thread. The Anthropic Administrator MCP is the right call here because it exposes Create Message and List Models as native tools—your team can prototype prompt chains, test model selection logic, and cache static system prompts all from the same chat where you're debugging the bot's behavior. The trade-off: you're writing raw API calls instead of using a pre-built integration, so this only makes sense if your workflow is already technical and you need that control. If you just want to ask Claude questions, use the standard Claude chat—this MCP is for teams treating Claude as a programmable component.

Use prompt caching when your team reruns the same context daily

A 3-person support team runs a daily digest where they paste 40 recent Slack threads into Claude and ask for sentiment analysis. They're hitting the same system prompt and thread format every morning, so costs add up. The Anthropic Administrator MCP's Prompt Caching tool is built for this: cache the static instruction block once, then only pay for the variable thread content on each run. You'll cut token costs by 60-80% if your reused prompt is over 1,000 tokens. The threshold: if your prompt changes every time or you only run this workflow twice a week, caching overhead isn't worth it. But for daily or hourly repetition with a stable template, this MCP pays for itself in week one.

When your team needs to A/B test Claude model versions

A 5-person content team is drafting blog posts and wants to compare Claude Opus output against Sonnet 4 for tone and length before committing to a model for production. The Anthropic Administrator MCP's List Models and Get Model tools let them query available versions, check context windows, and swap models mid-conversation without leaving Switchy. This is faster than toggling between the Anthropic console and your workspace. The limit: if you're just using one model and not experimenting, the standard Claude integration is simpler. But if your workflow involves testing multiple models or you're building a pipeline that needs to programmatically select the right model based on task type, this MCP gives you the API-level control you need.