Ollama

@ollama list all the models I have installed locally and show me which ones are best for code generation

Based on your local Ollama installation, you have 5 models available: **Code-optimized models:** - codellama:7b (3.8GB, modified 2 days ago) — Best for code completion and explanation - deepseek-coder:6.7b (3.8GB, modified 1 week ago) — Strong at multi-language code generation **General models:** - llama2:7b (3.8GB, modified 3 days ago) - mistral:7b (4.1GB, modified 1 week ago) - phi:2.7b (1.6GB, modified 2 weeks ago) For code generation tasks, I'd recommend codellama:7b or deepseek-coder:6.7b. Both are specifically trained on code and will give you better results than the general-purpose models.

This example demonstrates the MCP's ability to query your local Ollama instance and combine model metadata with reasoning. The MCP requires Ollama to be running locally (typically on localhost:11434). The actual model list depends entirely on what you've pulled via `ollama pull`, so responses will vary by user.

@ollama use the mistral model to generate a technical explanation of how JWT tokens work, keeping it under 100 words

I've generated this explanation using your local mistral:7b model: "JWT (JSON Web Token) is a compact, URL-safe authentication method consisting of three Base64-encoded parts: header (algorithm and token type), payload (claims like user ID and expiration), and signature (cryptographic verification). The server creates a JWT by signing the header and payload with a secret key. Clients include this token in subsequent requests. The server verifies the signature without database lookups, enabling stateless authentication. JWTs expire automatically and can't be revoked server-side without additional infrastructure." Generated in 2.3 seconds using 94 tokens.

This showcases direct text generation against a specific local model. The MCP sends your prompt to Ollama and returns the model's output. Response quality and speed depend on your hardware and which model you select. Larger models (13B+) produce better results but require more RAM and take longer to generate.

@ollama start a conversation with llama2 where I'm debugging a Python error, then ask it to explain why my list comprehension is failing

I've initiated a multi-turn chat session with your local llama2:7b model. Here's the conversation: **You:** I'm getting a TypeError when running this list comprehension: `squares = [x**2 for x in range(10) if x > 5]` **llama2:** That list comprehension syntax looks correct and shouldn't produce a TypeError. Can you share the exact error message? The code should create a list [36, 49, 64, 81] without issues. The TypeError likely comes from surrounding code — perhaps you're trying to use `squares` in an incompatible operation afterward? The conversation context is maintained, so you can continue debugging by sending follow-up messages through the same chat session.

This demonstrates the chat tool's ability to maintain conversation history across multiple turns, useful for iterative problem-solving. The MCP manages the message history locally, but context windows vary by model (typically 2048-4096 tokens for 7B models). Long conversations may truncate early messages, and the session resets if Ollama restarts.

When Ollama wins for rapid prototyping without API costs

A 3-person startup building a chatbot prototype needs to test 6 different prompt strategies across 4 model sizes before committing to a cloud provider. Ollama is the right call here: the team runs models locally, iterates on prompts with the raw-mode generate tool, and compares outputs from llama3, mistral, and gemma without burning through API credits. The chat-with-history tool lets them simulate multi-turn conversations, and the show-model-information tool surfaces which template each model uses. This works until the prototype needs production scale or the team grows past 5 people—at that point, the local-only constraint becomes a collaboration bottleneck. If you're pre-revenue and testing model behavior daily, Ollama keeps your burn rate low while you find product-market fit.

When local LLMs beat cloud latency for field teams

A 12-person field service company has technicians drafting customer emails from job sites with spotty internet. The support lead sets up Ollama on each laptop with a 7B parameter model and connects it through Switchy. Technicians use the OpenAI-compatible chat completion tool to draft responses to common issues—warranty claims, troubleshooting steps, follow-up scheduling—without waiting for cloud API calls or risking timeouts. The list-models tool lets the support lead verify which version each laptop is running during quarterly updates. This setup breaks down if the team needs real-time knowledge from a central CRM or if response quality demands a 70B+ model that won't fit on laptop hardware. If your team works offline more than 30% of the time and can live with smaller models, Ollama removes the connectivity dependency.

When on-premise inference is the only legal option

A 6-person legal ops team at a healthcare company needs to summarize patient intake forms but can't send PHI to third-party APIs under their BAA. They run Ollama on an on-premise server with a HIPAA-compliant model, then route requests through Switchy's workspace. The generate-text tool processes intake PDFs into structured summaries, and the get-version tool helps IT audit which Ollama build is deployed during quarterly compliance reviews. The team stays under the 8-tool limit because they only need generation and model introspection—no embeddings, no fine-tuning. This works as long as the server can handle the team's request volume and the model quality meets their accuracy bar. If you're in a regulated vertical where data residency isn't negotiable, Ollama is often the only MCP that keeps you compliant without custom infrastructure.