I Told AI What I Wanted to Run. It Found the Compute

Here’s a prompt that would have sounded unrealistic six months ago:

“I have a JSONL dataset and want to fine-tune Qwen3–8B with LoRA. Ocean MCP, find me an H200, estimate the cost, and help me get started.”

That’s it. No configuration files. No browsing provider dashboards. No manually comparing specs and pricing across different platforms. Just a plain language description of what you need and what you’re trying to do.

Ocean Network’s MCP server handled everything that came after that prompt. It discovered available H200 compute across the network, compared providers, walked through the fine-tuning workflow, and broke down the costs before a single job was launched.

That example isn’t hypothetical. It’s the kind of workflow ON MCP was built for, and understanding why it works requires understanding what the Model Context Protocol actually changes about how AI agents interact with infrastructure.

The Problem That ON MCP Solves

Most developers use AI as a thinking tool and treat infrastructure as a separate, manual problem.

You ask Claude to help you think through a fine-tuning approach. It gives you a solid answer. Then you go somewhere else entirely to actually find compute, compare pricing, figure out the right environment for your workload, and set everything up. The intelligence and the execution live in completely different places, connected only by you manually carrying information between them.

That separation creates friction that compounds at every step. You leave your thinking environment to find a GPU provider. You leave the provider to figure out the right configuration. You come back to your AI with questions, get answers, and go back to the infrastructure to implement them. The back and forth is constant, and none of it is doing the actual work you set out to do.

ON MCP removes that separation entirely. Your AI agent stays connected to Ocean Network’s compute infrastructure throughout the workflow. When you describe what you need, the agent doesn’t just advise you on what to do next. It goes out and does it, discovers providers, evaluates options, estimates costs, and walks you through execution without you manually bridging any gap.

What ON MCP Actually Is

ON MCP is a Model Context Protocol server that connects AI agents directly to Ocean Network. The Model Context Protocol itself is a standard that Anthropic built so AI agents could connect to external tools and services through a consistent interface rather than requiring a custom integration for every tool they need to use.

Think of it like a universal adapter. Instead of building a specific plugin every time you want your AI to interact with a new service, you build one MCP connection, and the agent gets access to everything on the other side of it.

ON MCP is Ocean Network’s implementation of that idea for compute infrastructure. Connect it to Claude, Cursor, ChatGPT, Gemini, VS Code, or GitHub Copilot and your agent immediately has access to roughly 80 tools covering the full surface of what Ocean Network can do. Discover compute providers, write and validate algorithms, run free or paid jobs, manage persistent storage, inspect node status and environments. All of it accessible through natural language without you writing a single integration.

The server is built peer-to-peer first. When it starts up, it joins Ocean Network’s libp2p network directly, which means it discovers and communicates with nodes across the network’s distributed hash table rather than depending on a single gateway. When your agent asks for compute, it resolves a node, negotiates an environment, runs the job near the data, and pulls back only the results. The infrastructure complexity occurs at the network layer and is invisible to you and your agent.

Back To The Fine-Tuning Example

The Qwen3–8B prompt illustrates exactly why this architecture matters for real workloads.

Fine-tuning a model with LoRA on a JSONL dataset isn’t a simple job. It involves specific GPU memory requirements, the right software environment, understanding how LoRA parameters affect training dynamics, estimating how long the job will run and what that costs, and making sure your dataset format is compatible with the training setup. Normally you’d be navigating all of that manually across multiple tools and dashboards.

With ON MCP, you describe the goal and the agent orchestrates the tools. It goes out across the network and discovers which nodes are advertising H200 capacity. It compares those providers across specs and pricing. It estimates the cost of running your specific workload against your dataset size and expected training duration. It walks you through the fine-tuning workflow so you understand what’s happening before you authorize anything. Only when you’re ready to proceed does it return the transaction for you to sign and broadcast yourself.

That last part matters and deserves its own attention.

Your Keys Stay With You

ON MCP never holds your funds and never signs anything on your behalf.

When a job requires payment, the server generates an unsigned transaction and returns it to you. You authorize it yourself through the Ocean Network Dashboard or your connected wallet. The agent does all the work of finding compute, negotiating environments, and preparing the job, but the moment money changes hands, that decision stays with you.

This is a meaningful design choice. Giving an AI agent autonomous access to your funds to execute infrastructure spending would create risk that most developers reasonably don’t want to take on. ON MCP separates the intelligence layer from the authorization layer. The agent can think, discover, plan, and prepare everything. Execution only happens when you explicitly approve it.

For quick experiments and testing, the server can use a temporary ephemeral key so you can run free compute jobs with no wallet connected at all. If you want to explore what ON MCP can do before committing any funds, that path is available from the first time you connect.

The Three Things Your Agent Can Do

Once connected, ON MCP gives your agent three distinct capabilities that cover the full workflow of working with Ocean Network.

• The first is running compute jobs. Your agent can find providers with the hardware you need, write and validate algorithms appropriate for your workload, and execute jobs against datasets you point it toward. Both free and paid compute are available. Free environments work for testing and smaller experiments. Paid environments give you access to serious GPU hardware, including H200s for demanding workloads like model fine-tuning.
• The second is publishing data. If you have datasets or algorithms you want to make available on Ocean Network, your agent can handle the publishing workflow. It manages the technical steps that would normally require navigating the Ocean CLI directly, accessible through plain language instructions instead.
• The third is node operation. Your agent can inspect node status across the network, check available compute environments, review job history, and monitor what’s running. For anyone operating nodes on Ocean Network, this means an AI agent can help manage and monitor infrastructure conversationally rather than through manual dashboard navigation.

How This Compares To Other Ocean Tools

ON MCP sits alongside Ocean’s other interfaces rather than replacing them, and the distinction is worth understanding.

The Ocean Orchestrator is a VS Code extension built for developers who want to author and launch jobs interactively. You’re hands-on with each step, making decisions manually with full visibility into what’s happening at every point. This is the right tool when you want direct control and are comfortable with the workflow.

The Ocean CLI drives the network from your terminal. It gives developers direct command-line access to everything Ocean Network can do. If you prefer scripting workflows or want to integrate Ocean into existing automation, the CLI is the appropriate interface.

ON MCP is the agent interface. You describe what you want to accomplish and the AI figures out which of those roughly 80 tools to call, in what order, with what parameters. The orchestration happens automatically. You stay at the level of goals and outcomes rather than individual tool calls and configuration decisions.

The right tool depends on how you prefer to work. Hands-on and manual, the Orchestrator. Terminal and scriptable, the CLI. Natural language and agent-driven, ON MCP.

What This Means For How Work Gets Done

The Qwen3–8B fine-tuning example points toward something more significant than a convenient workflow shortcut.

When you can describe a complex infrastructure task in plain language and have an agent handle discovery, comparison, cost estimation, workflow guidance, and job preparation automatically, the distance between having an idea and executing it collapses considerably. You’re not managing the infrastructure anymore. You’re working at the level of what you’re trying to build.

For solo developers and researchers who don’t have dedicated infrastructure teams, this matters a lot. The cognitive overhead of managing compute infrastructure is real, and it compounds. Every decision about which provider to use, what specs to choose, how to estimate costs, and how to configure the environment is time and attention that isn’t going toward the actual work.

ON MCP moves that overhead to the agent layer. You focus on the problem. The agent handles the infrastructure decisions that support solving it.

That’s not a minor convenience. It’s a different relationship between intelligence and execution that changes how quickly things actually get built.

Visit Ocean Network here