Raw: https://promptreports.ai/docs/quickstart.md

# Quickstart

Three steps, once the app and the runner ship: connect the products, get an agent token, run your first workflow.

## 1. Connect the products

Sign in at app.promptreports.ai with PromptReports ID and open Connections. Click Connect next to each product. Each one runs PromptCompute's connect flow and returns a key that can only ask for that product's actions.

Start with RepoOps and PromptCompute: most workflows use both.

## 2. Get an agent token

Register your agent on the Agents page, or from code:

```
POST https://api.promptcompute.ai/v1/agents          # built, needs agents:write
GET  /v1/agents/{id}/key-challenge                   # sign the nonce with the agent key
POST /v1/agents/{id}/keys
POST /v1/agents/{id}/versions                        # signed manifest: code, model, prompt hash
```

Then connect your MCP client. It signs in through the browser; no key goes in the config:

```
{ "mcpServers": {
  "promptreports": { "type": "http", "url": "https://mcp.promptreports.ai/mcp" },
  "promptcompute": { "type": "http", "url": "https://mcp.promptcompute.ai/mcp" }
} }
```

promptcompute is built. promptreports is Proposed. Clients identify with a Client ID Metadata Document.

## 3. Run your first workflow

Start with a dry run of [Launch with proof](https://promptreports.ai/use-cases/launch-with-proof/). A dry run will count no workflow request, and it shows which steps would record missing evidence.

```
list_workflows()
start_run({ workflow: "launch-with-proof",
            inputs: { release_tag: "v2.8.0" }, dry_run: true })
get_run({ run_id: "R-2041" })
# { status: "done", counted: false, steps: [...] }
```

Then run it for real. The run stops at the approval; a person decides in the approvals inbox, and the run resumes.