Get started in 5 minutes

This walks you through signing up, connecting a platform, registering an agent, and watching a real action flow through AgentValet end-to-end.

1. Sign up

Go to app.agentvalet.ai  and create an account. You can sign up with email or GitHub.

On the Free plan you’re first asked to pick a plan (Continue with Free is fine). Then you land on the dashboard with a short tour and a setup guide. You can follow them or skip them — this guide covers the same ground.

2. Connect a platform

Open Platforms from the sidebar. Pick something low-stakes for your first connection — a Slack workspace you can experiment in, a personal Gmail, a Linear workspace you don’t share with anyone.

Click Connect. For OAuth platforms you’ll go through the provider’s normal authorisation flow. For API-key platforms you’ll paste a key. Either way, AgentValet encrypts what you give it and stores it in a vault — you can disconnect any time.

3. Register your first agent

Pick the agent runtime you actually use.

Option A — Claude Desktop

1. Download the AgentValet extension  (agentvalet-latest.mcpb).

   • On macOS and most Windows installs: drag the .mcpb file into Claude Desktop’s extensions window.
   • On Windows 11 / Claude Desktop from the Microsoft Store: drag-and-drop is sometimes blocked. Use Settings → Extensions → Advanced settings → Install extension and pick the .mcpb from disk.

   

2. In the dashboard, open Agents → Register new agent, name it (say, My MacBook), pick Claude Desktop (MCPB) and click Register.

3. The dashboard shows four values: Agent ID, Owner ID, Proxy URL and Agent Private Key (base64). The key is shown once. Paste them into the install dialog Claude Desktop shows for the extension, then restart Claude Desktop.

Option B — CLI (Claude Code, Cursor, Codex, scripts, codebases)

In your project directory:

npx @agentvalet/register

The CLI opens your browser to sign in, then generates a keypair on your machine (CLI 1.7.0 and later), registers the agent under your account, writes the private key to .agentvalet/agent.key in the project directory, and wires up the MCP config for the tools it detects. See Register via the CLI for exactly what it writes.

Either way, the agent is active straight away and starts with no platform access. You grant that in step 5.

4. Confirm it’s wired up

Before granting any scopes, ask the agent to call list_platforms. It returns the platforms approved for this agent — initially an empty list, and that’s fine. What matters is that the call succeeds.

This one call proves three things at once:

• your agent’s private key is valid and signs correctly
• the AgentValet proxy is reachable
• the agent record exists and is marked active

If list_platforms fails here, fix it before going further — granting scopes against a broken setup just makes the failure harder to diagnose later. The error message usually points at which of the three pieces is wrong.

5. Grant the agent some scopes

In Agents → [your new agent], click + Grant access. Pick the platform you connected in step 2 and select the scopes you want this agent to have.

Each grant carries a require approval toggle that covers every call on that grant, reads included. For per-action control, set actions to Ask first in a policy. A few rules of thumb:

• Read scopes — usually safe to leave auto-approved.
• Write scopes — start with approval on. Drop it once you’ve watched the agent for a while.
• Destructive scopes (delete, send, charge, archive) — keep approval on. The system will let you turn it off; you shouldn’t.

See Scopes & permissions for the full model.

6. Make a call

Ask your agent to do something using a scope you just granted. For example: “Post a hello message to my #test channel in Slack.”

What happens behind the scenes:

1. The agent signs a request with its private key and sends it to AgentValet.
2. AgentValet verifies the signature, checks the agent has chat:write for Slack, decrypts your Slack credential, and forwards the call to Slack.
3. Slack returns success. AgentValet logs the action — agent, platform, scope, timestamp — to your audit log.

Check Audit in the sidebar. Your call is there. Nothing the agent did to your Slack is unobservable.

7. Try something that needs approval

Ask the agent to do something destructive — delete a Slack message, send an email, archive a channel. This time, instead of acting, it’ll come back with: “That’s queued for approval.”

Open Approvals. Approve or deny. If you approve, the original action executes and you see the result. If you deny, nothing happens and the denial is recorded in the audit log.

That’s the whole loop. Connect → register → grant → call → audit (or approve). Everything else in the product is a refinement of these five steps.

Prefer a browser and ten minutes?

Open the lab in a GitHub Codespace: labs/01-governed-loop . Nothing installs on your machine and there is no third-party OAuth: the upstream is a mock MCP server the broker serves for exactly this purpose. You register an agent, make a governed read, get a call held for your approval, see deny-by-default, revoke the agent, and verify the signed receipt against the published key set.

Prefer to start from working code?

Every step above is also a runnable example: github.com/AgentValet/agentvalet-examples .

Pick the one matching how you build — Claude Code (no code at all), a plain Node script, Python inside a LangGraph node, or a CrewAI crew where each member holds its own attenuated identity. Each runs the same loop end to end, and they’re tested nightly against the published packages, so they work today rather than on the day they were written.

Next steps

• How approval works in detail
• Plans & pricing
• Install paths compared