Install from zero

This guide is for a person who has never built a Relevance AI Tool. The goal is to create a safe development copy of the RunOnProof Tool and connect it to a test agent without paying RunOnProof or calling a paid production verification.

Before you start

You need:

• access to a Relevance AI project you are allowed to modify;
• access to this repository;
• the files authoring-manifest.json, schemas/tool-input.schema.json, runtime.mjs, and SYSTEM-PROMPT.md;
• about 20–30 minutes for a first installation.

You do not need a wallet, payer key, payment signature, or RunOnProof secret for the safe synthetic setup. Do not create or paste payment credentials into Relevance AI.

Know the three environments

• Synthetic test uses repository fixtures. It does not contact a real source and costs 0.00 USDC.
• Staging is a RunOnProof engineering environment. Use it only when a documented staging URL and authorization exist for the test.
• Production is https://api.runonproof.com. A paid POST can return HTTP 402 and may lead to a charge only through a separately authorized x402 payment client. This integration must stop before payment.

If you are unsure which environment you are using, stop and inspect the URL before pressing Run.

Installation sequence

1. Open Relevance AI and sign in.
2. Open a development project. Do not begin in a shared production project.
3. Open Tools.
4. Select New Tool and choose Start from scratch. Button labels can change; if the interface differs, choose the option that creates an empty custom Tool rather than a template.
5. Build the Tool by following Create the Tool.
6. Test the Tool with the synthetic cases before connecting it to an agent.
7. Open Agents, create a blank agent, and follow Create and connect the agent.
8. Confirm the Tool execution policy says Requires approval.
9. Run the seven cases in Fixtures and expected results.
10. Record the Tool version, agent version, project, test date, and tester. Do not record secrets.

How to know installation succeeded

The installation is successful only when all of these are true:

• the Tool appears in the project Tool list;
• the agent lists the Tool under its connected resources or tools;
• the Tool requires approval before execution;
• there is no wallet, payer key, PAYMENT-SIGNATURE, cookie, or reusable payment proof in the Tool;
• the payment-required fixture returns REVIEW and shows the product and price;
• conditioned capabilities stop before HTTP;
• a source outage does not become PROCEED;
• replay is described as no-new-source-call and no-new-charge;
• the test history shows no unintended production POST or settlement.

Do not do these things

• Do not paste an API key into a prompt, fixture, URL, screenshot, or chat.
• Do not enable automatic retry for a paid POST.
• Do not turn HTTP 402 into a successful result.
• Do not publish the Tool or agent publicly as part of installation.
• Do not claim a Relevance AI partnership.
• Do not use a real company as the first test when a synthetic fixture can prove the contract.

Existing installations

Before creating a Tool, check your own project for an existing version. Avoid duplicates; recreate only for isolated training, disaster recovery, or a deliberately versioned replacement. This public repository does not expose RunOnProof's private workspace identifiers or links.