keyring, by Stillloom
Let agents log in.
Never hand them the password.
Keyring is a broker for customer logins, sessions and one-time codes. Your agents and bots ask for a session. Keyring signs in, handles the code, and stops them from locking each other out.
Alpha. In development and not open yet. This page describes what we are building.
Diagram. Three agents, A, B and C, ask Keyring for the same account. A gets the lease with fence number 41 while B and C wait. An SMS code arrives and is routed to A's lease only. Keyring returns a session to A with no password in it. A releases the lease, the fence moves to 42, and B takes the lease.
01 / The problem
Logins are where agent workflows quietly break.
- 1
Two workers sign in as the same user at once. The portal logs one out, or locks the account.
- 2
A code lands on a shared phone or inbox, and the wrong worker uses it.
- 3
A customer password expires, and nobody knows until the sync has failed for days.
Every team ends up writing this plumbing itself. Keyring is that plumbing, built once.
02 / How it works
Ask for a session. Get a session.
- 01
Lease.
A worker asks for a login. Keyring lets one worker at a time sign in to an account and gives it a fence number. A worker that has lost its lease is refused.
- 02
Sign in.
Keyring runs the portal's login in its own browser, with the stored password. The worker never sees it.
- 03
Handle the code.
If the portal asks for a code, Keyring works it out from the TOTP seed, or catches the SMS or email and matches it to the one login that is waiting for it.
- 04
Keep alive.
The session is reused and refreshed, so the next run skips the login.
Every step is written to an audit log.
03 / Agent first
Built for agents to call, not only for people to set up.
Keyring has an MCP server next to its HTTP API. One call, get_session, takes the lease, signs in and waits for the code. When something goes wrong, the error says what to do next: wait, try again later, or ask a human.
MCP
The call.
{ "tool": "get_session",
"arguments": { "account": "acct_7Q2", "idempotency_key": "job-8841-session" } }The result. A session handle, a lease and a fence. No password.
{ "lease_id": "lse_51", "fence": 41, "expires_at": "2026-11-20T09:15:00Z",
"heartbeat_every_s": 30, "session_handle": "wss://..." }An error, with the next step spelled out.
{ "code": "lease_busy", "message": "Another worker holds acct_7Q2 until 09:15:00Z.",
"retryable": true, "retry_after_s": 40,
"next_action": { "tool": "get_session", "args": { "account": "acct_7Q2", "wait_s": 60 } } }TypeScript
import { Keyring } from "@stillloom/keyring";
const keyring = new Keyring({ url: process.env.KEYRING_URL, apiKey: process.env.KEYRING_API_KEY });
await keyring.withLease("acct_7Q2", async ({ sessionHandle, signal }) => {
await runAgent(sessionHandle, signal);
});withLease sends heartbeats, passes the fence on every call, and releases the lease when your code is done. If the lease is lost, signal aborts.
CLI
keyring session get --account acct_7Q2 --json
keyring session release --lease lse_51 --fence 41 --outcome ok --jsonEvery command takes --json and returns the same result and the same errors as the API.
Preview. These names and arguments follow the draft contract and may change before 1.0. Python and Go SDKs are planned.
04 / Controls
The boring parts, done carefully.
One login at a time.
Each account has a lease with a fence number. The lease runs out on the broker's clock, not the worker's.
Lockout budgets.
A login attempt is counted when it starts, so a crash still counts. One try is always held back for a human.
Session reuse.
Sessions are kept warm and reused, so most runs skip the login.
Code relay.
TOTP, SMS and email. A code goes only to the login that is waiting for it. A shared phone or inbox is held to one login at a time.
Health checks.
Know a login is broken before a customer or a failed job tells you.
Audit log.
Insert-only, and hash chained per tenant. It never holds secret values.
05 / Plugins
Your portal, your vault, your language.
Keyring is small at the core and grows through plugins. A plugin can be written in any language and runs in its own process.
Credential store.
Holds the sealed secrets.
Key management.
Wraps and unwraps the keys.
OTP channel.
Gets one-time codes.
Browser runtime.
Gives a login flow a browser.
Portal login flow.
Login, check, keep alive and log out for one portal.
Audit sink.
Copies audit events out.
- A portal flow is one file: login, check, keep alive, log out.
- A plugin gets only the secret fields it declares, for one lease at a time, and can reach only the hosts it lists.
- A test kit runs each plugin against a fake portal with codes, lockouts, expiry and CAPTCHA.
06 / Security
Agents get a session, not the password.
By default Keyring signs in for the worker. If a job truly needs the password, an account can be switched to allow one short reveal per lease, and every reveal is logged. Everything else stays in the vault.
How it is designed
- Each secret is encrypted with its own key, under a managed key service, and tied to its tenant, account and field.
- Cookies and sessions are treated as secrets.
- TOTP seeds never leave the broker.
- Tenants are kept apart in the database.
- Each portal flow records a review of the portal's terms. Unreviewed portals are refused.
What we do not claim
Keyring is in alpha. It has had no outside security audit and it holds no certificate of any kind. If a portal shows health or financial data, you still need your own agreements and your customers' permission. Keyring is a tool, not a certificate.
07 / Where this stands
Early, and open about it.
Keyring is made by Stillloom and is in alpha. Today there is a written design and the first code. There is nothing to install yet. The first code channels will be TOTP, SMS and email. The API, the plugin interface and the package names may change before 1.0.
Join the waitlist and we will write when there is something to try.
Want it early?
Leave your email. We will write once, when access opens.