authyouragent (Python)

SDK for Auth Your Agent: let AI agents act for a person on websites, with the person's approval on their phone.

• AgentClient: an agent asks for access, then calls sites with a short-lived pass and a fresh DPoP proof (RFC 9449) on every request.
• SiteVerifier: a website checks each call in one line.
• Browser vault: the agent's browser runs in a container it cannot read into. When a site asks for a password, a CAPTCHA or 2FA, the owner takes over from their phone; clicks that commit something wait for the owner's approval.

Python 3.9+. Depends on httpx, pyjwt, cryptography.

Full documentation: https://authyouragent.com/docs/developers/quickstart

Install

You need Python 3.9+ and Docker.

pip install "authyouragent[mcp]"
authyouragent vault up --agent-id ag_xxxxx --key /path/to/agent-key.pem

vault up starts the browser vault: the browser your agent uses, in a container on your machine. The agent drives it through the vault and never gets its cookies. It prints the env block for your MCP client. If you skip it, the MCP server starts the vault itself the first time the agent needs the browser (AYA_VAULT_AUTOSTART=0 turns that off). See vault/README.md for what protects what, and the known limits.

For just the SDK (no browser, no MCP): pip install authyouragent.

MCP server

The authyouragent-mcp command starts an MCP server with twenty-two tools:

• Browser: navigate, read_page (the page's text plus a numbered list of links, buttons, fields and dropdowns), and click, type_text, select_option, press_key by that number. scroll, go_back, wait_for, screenshot. Only public websites open. Anything that submits a form, or whose button says create, send, save, delete, pay and the like, waits for your approval on your phone.
• Approval modes (in the app, per agent and site): Ask, Smart or Allow (one site at a time). Smart lets changes you can undo (save, filter, add to cart) through without asking; money, deleting, posting and account changes always ask. Rules only, no AI. The approval card shows the amount and item read from the page, and the vault will not click if the amount changed after you approved.
• Trusted sites: authyouragent vault trust example.com --no-approvals stops per-click approvals on a site you use often; authyouragent vault trust 192.168.1.20:8123 --private lets the browser open one host on your own network. Each entry asks your phone once per session before it applies.
• check_login_wall: is the page asking for a password, a code or a sign-in approval?
• list_secrets / fill_secret: fill a username, password or authenticator code from your own Bitwarden or Vaultwarden (only the items in its "Auth Your Agent" folder). The agent never sees the value; it fills only on the item's own site and only the right kind of field.
• request_takeover / wait_for_takeover: you take over the browser from your phone. The agent is disconnected until you finish, and the vault hands back by itself once you have signed in.
• request_approval: ask you to approve an action the vault cannot see.
• notify_owner: a one-way note to your phone, e.g. "done" or "stuck". Nothing to approve.
• submit_plan / plan_status: for a scheduled task. The agent lists what it will do, when and in what order; you pre-approve all steps, some or none, once, on your phone. In the window, an action that matches a pre-approved step exactly (address, button words, the exact text it posts) goes through without a card. A step you do not answer in time is skipped with the steps chained to it; the rest of the plan carries on. Payments and security changes always ask. For a task that repeats, the agent asks for an exact number of runs (runs, every = day or week, all within 30 days); one approval covers exactly those runs and no more.
• end_session: sign out of every site used, then destroy the browser profile. Always called at the end. The vault then sends you its own summary of the session, whether or not the agent sent a note.
• check_agent_status: is the agent still authorized? You can revoke it at any time.
• report_site: report a site where take over did not work.

Configuration

AYA_AGENT_ID=ag_xxxxx
AYA_KEY_FILE=/path/to/agent-key.pem
AYA_VAULT_URL=http://127.0.0.1:7801
AYA_VAULT_TOKEN_FILE=~/.authyouragent/vault/token

Hermes

mcp:
  authyouragent:
    enabled: true
    command: authyouragent-mcp
    env:
      AYA_AGENT_ID: "ag_xxxxx"
      AYA_KEY_FILE: "/path/to/agent-key.pem"
      AYA_VAULT_URL: "http://127.0.0.1:7801"
      AYA_VAULT_TOKEN_FILE: "~/.authyouragent/vault/token"

Claude Desktop / Cursor

Claude Desktop: claude_desktop_config.json. Cursor: ~/.cursor/mcp.json.

{
  "mcpServers": {
    "authyouragent": {
      "command": "authyouragent-mcp",
      "env": {
        "AYA_AGENT_ID": "ag_xxxxx",
        "AYA_KEY_FILE": "/path/to/agent-key.pem",
        "AYA_VAULT_URL": "http://127.0.0.1:7801",
        "AYA_VAULT_TOKEN_FILE": "~/.authyouragent/vault/token"
      }
    }
  }
}

Agent

python -m authyouragent keygen --name "Job-search assistant"

Register the printed jwk in the app (Agents -> Add an agent). Keep privkey_pem on the agent's machine.

Gate a script or CI step on your phone. The command exits 0 only if you approve:

export AYA_AGENT_ID=ag_xxxxx AYA_KEY_FILE=/path/to/agent-key.pem
authyouragent approve npmjs.com publish && npm publish

Exit codes: 0 approved, 1 denied, 2 could not ask (key, revoked, network), 3 no answer in time. From Python: agent.request_approval("npmjs.com", "publish").

Tell your phone when an unattended job ends (one-way, nothing to approve):

make deploy; authyouragent notify "deploy finished with exit code $?"

Exit codes: 0 delivered, 1 recorded but no device has notifications on, 2 could not send. From Python: agent.notify("done").

from authyouragent import AgentClient

agent = AgentClient(base_url="https://authyouragent.com", agent_id="ag_…",
                    privkey_pem=open("agent-key.pem").read())
agent.ensure_grant("jobs.example.com", scopes=["list", "apply"])   # phone approval
r = agent.request("GET", "https://jobs.example.com/api/jobs")
r = agent.request("POST", "https://jobs.example.com/api/jobs/j1/apply", stepup_action="apply")

Sites with a Sign in with Auth Your Agent button (OpenID Connect): the vault handles it on its own when the agent clicks the button. Without the vault:

back = agent.oidc_signin(authorize_url)   # the site's /oidc/authorize link; waits for the phone if needed
http.get(back)                            # finish on the site's own callback

Site

from authyouragent import SiteVerifier, AuthError

verifier = SiteVerifier("https://authyouragent.com", expected_audience="jobs.example.com",
                        public_base_url="https://jobs.example.com")

@app.get("/api/jobs")
def jobs(request: Request):
    try:
        auth = verifier.verify(request)
    except AuthError as e:
        raise HTTPException(401, {"error": e.error})
    ...

See examples/ for a complete agent and a complete FastAPI site.

Licence

MIT