OpenAI

@OpenAI list all fine-tuning jobs from the last 30 days and show me which ones completed successfully

Found 4 fine-tuning jobs in the last 30 days. Two completed successfully: 'ft-abc123' (GPT-3.5-turbo fine-tune on customer support data, finished Jan 15) and 'ft-def456' (GPT-4 fine-tune on legal documents, finished Jan 22). One job 'ft-ghi789' failed due to insufficient training examples. One job 'ft-jkl012' is still running (82% complete). The successful models are now available in your account and can be called using their fine-tune IDs.

This example shows read-only monitoring of fine-tuning operations. Requires an API key with access to your organization's fine-tuning history. The MCP returns job metadata but cannot access the underlying training data or model weights. Useful for tracking training pipeline status without leaving your workspace.

@OpenAI create a new assistant named 'Code Reviewer' that uses GPT-4 and has code interpreter enabled

Created assistant 'Code Reviewer' with ID 'asst_xyz789'. Configuration: model set to gpt-4-turbo-preview, code interpreter tool enabled, no file search or function calling. The assistant is now active and can be used in threads. You'll need this assistant ID to create threads and send messages. Note that using this assistant will incur GPT-4 API costs per message, and code interpreter executions count toward your usage limits.

This example demonstrates creating a persistent assistant resource. The assistant remains in your OpenAI account until explicitly deleted, and each message sent to it incurs API costs. Be aware that assistants with code interpreter can execute Python code in a sandboxed environment—review security implications before using with untrusted input.

@OpenAI check what models are available in my account, then recommend which one I should use for summarizing technical documentation

Your account has access to 12 models including GPT-4-turbo, GPT-4, GPT-3.5-turbo, and several legacy versions. For summarizing technical documentation, I recommend GPT-4-turbo-preview: it has a 128k token context window (can process ~300 pages in one request), strong technical comprehension, and costs $0.01 per 1k input tokens. If cost is a priority, GPT-3.5-turbo-16k works well for shorter docs at $0.001 per 1k tokens but may miss nuanced technical details. Avoid the base GPT-4 model—it has only an 8k context window.

This example pairs the MCP's list-models capability with the AI's reasoning to provide contextualized recommendations. The model availability and pricing shown here are illustrative—your actual account access depends on your OpenAI organization tier and any custom rate limits. Always verify current pricing before committing to high-volume usage.

When this MCP wins for rapid assistant iteration

A 3-person AI product team is building a customer-facing chatbot with multiple specialized assistants (billing, onboarding, technical support). They need to spin up new assistants, test conversation flows, and delete failed experiments daily. The OpenAI MCP is the right call here because it exposes the full Assistants API lifecycle—create, message, delete—without leaving Switchy. The team can iterate on instructions and tool configs in shared chats, then push changes via the MCP in under a minute. The threshold: if you're managing more than 10 assistants or need version control on assistant configs, you'll outgrow this MCP and want a proper CI/CD pipeline. For early-stage prototyping with a handful of assistants, this MCP cuts iteration time in half.

When this MCP barely justifies the setup cost

A solo ML engineer is running weekly fine-tuning jobs on GPT-3.5 for a niche legal document classifier. She wants to check job status and list completed runs without opening the OpenAI dashboard. The OpenAI MCP technically works for this—list fine-tunes and list models cover the read-only monitoring—but it's borderline overkill. The MCP requires an API key in Switchy's settings, and the payoff is just avoiding a browser tab. If she's already in Switchy for other integrations (Linear, GitHub), the marginal value is there. If this is the only OpenAI task she does in Switchy, she's better off bookmarking the dashboard. The buying threshold is whether you're doing at least two other OpenAI workflows in the same workspace.

When this MCP handles the unglamorous file hygiene work

A 6-person support team uploads hundreds of customer transcripts and PDFs to OpenAI for retrieval-augmented generation in their internal knowledge assistant. Every month, they need to audit and delete outdated files to stay under quota. The OpenAI MCP is the right tool because it exposes list files and delete file in a shared workspace where the whole team can see what's being purged. One person runs the list, the team flags files for deletion in a Switchy thread, then someone executes the batch delete—all without juggling API keys or writing throwaway scripts. The trade-off: if your file count exceeds 500, the list tool gets slow and you'll want a dedicated admin script. For teams under that threshold doing quarterly or monthly cleanups, this MCP turns a 2-hour task into a 20-minute team exercise.