This is what working with heel looks like day to day: what you set up once, what your agent sees, and what you see when it needs you. Running heel covers installing and configuring it.
Register your passkey.heel user enroll owner prints a one-time link; opening it registers a passkey on that device.
On localhost, that's this computer: you approve here, with Touch ID, say.
To approve from your phone, give the approval page an HTTPS address first, and set it as public_url before you enroll, because a passkey is tied to the address. Tailscale (tailscale serve --bg 7301), a Cloudflare tunnel or a server with TLS all work. Only port 7301 needs to be reachable. Then open the enrollment link on your phone, add the page to your home screen, and allow notifications. Approve from your phone has the steps.
Connect the accounts your agent should use. Each one gets an alias, and the agent only ever sees the alias.
gh auth token | heel connect github --as github --credential-stdinheel connect gmail --as gmail-work # then sign in from the Connections pageheel connect mcp --as notion --url https://mcp.notion.com/mcp --oauth
Any MCP server, any HTTP API with an OpenAPI file, and any CLI can go behind heel. heel --help lists the forms.
Write your rules in $HEEL_HOME/rules/owner.yaml (Rules). If you're unsure where to start, run heel serve --watch for a while. Every call runs, and the Activity page shows what each rule would have done.
Add the agent.heel agent add my-agent prints its heel token, once, and the MCP URL. That token is the only credential the agent gets.
The agent now has one MCP server with every connected account's tools, named by alias (gmail-work__send, github__merge). Or give it one account at a time at /mcp/<alias>.
On Linux,/mcp without ?tools=direct is code mode: one run_code tool, where the agent writes a short Python script against typed modules (gmail_work.search(...)).
Claude, ChatGPT or Grok connectors sign in with OAuth. Set agent_public_url (Running) and add heel as a custom connector.
CLIs: the agent runs them as heel run -- gh pr merge 123, with HEEL_TOKEN and HEEL_URL set.
HTTP APIs: the agent calls /http/<alias>/<path> with its heel token.
You get a notification on your phone (web push), or by email or Slack, if heel has an HTTPS address (setting one up). It says which agent wants to do what, and links to heel. Replying to it approves nothing. Without notifications, waiting requests are listed on the approval page.
You open the request. The page shows the account, the tool, the rule that asked, the counts, and the full arguments: every recipient and the whole body, exactly what will run. If you've turned on the optional hints, they appear under the call ("asks someone for money", "looks routine for this agent"). Without them, everything else here works the same.
You decide, with your passkey:
Approve once: this exact call runs the next time the agent sends it, once.
Approve as standing: also allow this exact call, or everything this rule asks about, for a while (an hour up to 90 days), so you aren't asked again. Standing approvals are listed on the Standing approvals page, where you can revoke them.
Deny: the agent gets "denied" when it re-sends.
The agent re-sends and the call runs, checked against your rules again. A request nobody answers expires after 24 hours.
Tap Pause on the approval page, or run heel pause --agent my-agent. You can pause one agent, one account, or everything. Every call in scope stops at once, even one you already approved. Resuming needs your passkey, and nothing called during the pause is replayed.
Activity shows every call, with its decision and the rule that made it, as they happen. It also suggests rules from what got asked or denied often.
Grants and standing approvals list what you've allowed beyond your rules, and let you revoke any of it.
Modes let a new rule, or a new agent, run in watch mode first: heel logs what the rule would have done, then you switch it to enforce.
Rules: edit the file and heel reloads it. Run heel rules check and heel rules test before you rely on a change. Editing a rule ends the grants made from it, so widening a rule asks again.