Quickstart

Before you begin

You need three things:

• Your model project on your machine, in the state it runs in today.
• A coding agent you use in that project: Claude Code, Codex, or any agent that can read a skill file.
• A RunLocal account. Access is by invitation. If you don’t have one, book a call.

You don’t need an API key, and you don’t need to convert or export your model first. Your agent works from the source as it is.

You run steps 1 and 2. Your agent does the work in steps 3 to 6, and stops for your answer each time. You finish in the workspace at step 7.

1. Open your project and install the skill

1. In a terminal, change into your model project: cd path/to/your-model-project
2. Start your coding agent from that folder. It reads this project’s source, and it creates the project’s .runlocal/ folder here.
3. Select your agent below and run the commands shown.
4. Start a new session in the same folder, so the skill loads.

/plugin marketplace add neuralize-ai/runlocal-agent-plugins

/plugin install runlocal@runlocal

codex plugin marketplace add neuralize-ai/runlocal-agent-plugins

codex plugin add runlocal@runlocal

curl -fsSL https://www.runlocal.ai/skill.md -o SKILL.md

The skill tells your agent how to prepare and submit the package. Reading the full skill is optional.

2. Send your agent the task

If you installed the plugin, begin your message with the skill command shown above. If you pasted the skill, continue in the same chat.

Click Copy prompt, paste it into your message, add your model and your goal, and send.

Use the RunLocal skill to prepare and submit a Request Package for this model. Keep private weights local. Before remote validation or upload, list every entrypoint, function, operator, example, and file the package will send, and ask me to confirm that list.

For example, add: “Use the model in models/perception.py. I want lower latency on Jetson Orin NX.”

3. Answer your agent’s questions

Your agent reads your code first, then asks about what the code cannot tell it. Have these answers ready:

• Which model, if your project holds more than one variant.
• Which function you call in production, and where it starts and ends.
• Typical input sizes, and which of them vary between calls.
• Your target hardware, such as a phone, a laptop, or Jetson Orin NX.
• What you want improved: latency, memory, power, or size.

Answer in your own words. If you don’t know something, say so: an open gap is recorded in the request instead of guessed.

4. Review what will be sent

Your agent shows you a preview before anything leaves your machine: the scope, the files and their sizes, the values it keeps local, and the open gaps. Read it and check that:

• It covers the model or part of the stack you chose.
• You recognize every file and every piece of sample data.
• Trained weights, calibration results, and private samples stay local.
• Nothing in a name, note, or graph discloses more than you intend.

Ask about anything unfamiliar and request changes. When the preview is right, tell your agent to continue.

5. Sign in when your agent asks

Skip this step if your agent is already signed in.

1. Open the sign-in link your agent gives you.
2. Sign in with your invited RunLocal account.
3. If asked, enter the code your agent shows.
4. Complete sign-in, then return to your agent.

Use the same account you will open the workspace with at step 7. Your agent keeps the session in memory only.

6. Let your agent validate and upload

Your agent now checks the package and submits it. Watch for three things:

1. Validation problems. Have your agent fix them before it uploads. An unknown value is recorded, not erased.
2. The work menu. Before submitting, your agent shows which kinds of work your information permits, and what each other kind would depend on. Adding more is your choice, not a requirement.
3. The result. On success your agent gives you the request identity and a workspace link.

Keep the .runlocal/ folder your agent created in your project, and commit it. It connects this request and later results to your code. It is not the ~/.runlocal/ folder in your home folder, which holds only API keys.

7. Check the upload in your workspace

Do this yourself, in a browser. An accepted upload means the package was well formed. It is not proof that it describes the right model, or that nothing private came along. You are the one who can tell.

1. Open https://www.runlocal.ai/app and sign in with the same account you used at step 5.
2. The workspace opens on Requests, newest first. Your submission is at the top. Open it.
3. Step through the tabs and confirm what you see: Summary for the scope and goal, Model for the graph, files, and entry points, Weights for names, shapes, and data types with no trained values, Examples for the input sizes and sample calls, Knowledge for the gaps left open, and Raw files for exactly what was uploaded.
4. Open Models to see one row per model with its latest revision, so you can tell which revision is current.

If something is wrong, tell your agent what to correct and have it submit again. Every revision is kept, so you can compare them. To remove an upload, delete the request in the workspace: the stored package is removed and its quota released.

What happens next

RunLocal assesses your request and proposes performance targets with a fixed price and an estimated timeline for each. You approve a target and quote before optimization starts. You only pay for an agreed target that is achieved.