Security
This page describes how the hosted product at teamshared.com handles sign-in, organization data, agent runs, and secrets. It describes the software. It is not an audit report. To report a vulnerability, see the end of this page.
Sign-in
You sign in with an email address and a one-time code. There is no password. We email a six-digit code. We store a hash of the code, not the code, and it works once. It expires after 15 minutes. Asking for a new code replaces the previous one. After five wrong guesses that address is locked for the same 15 minutes, and asking again does not send a new code or clear the count. A code can be sent three times a minute for an email address, and checked five times a minute. Checking a code is refused when the limiter cannot be reached. The first sign-in creates a private organization for that email.
Sessions and agent keys
A correct code sets an HttpOnly cookie named ts_session. It is a signed token for that person and one organization. Page scripts cannot read it. When the site is served over HTTPS, the cookie is marked Secure. It is not sent on a form post from another site. It lasts 30 days and is renewed while you use the console. Signing out revokes that sign-in, so other copies of the cookie stop working. Console form posts are checked against a second cookie tied to the same sign-in.
Agents connect over remote MCP. OAuth clients use the same emailed code. A harness that cannot do OAuth uses a tsk_ API key from API Keys (sign-in required). The raw key is shown once and stored as a one-way hash. Each key belongs to one organization.
Organizations
Memory, work, and files are stored per organization. The application connects as a database role that cannot bypass row-level security. Those tables return only the rows for the organization on that request, and they return nothing when that organization is not set. A personal organization stays with its owner. A company organization lets a member, and an agent that member connected with the email code, open another organization they belong to. That is the default, and the organization can turn it off. An API key, a dedicated agent, and an agent run stay in the organization they were issued for.
Agent runs
A TeamShared Cloud Agent run runs the model in a separate, sandboxed process. That process does not receive the database credentials, the session secret, or the run's API key. TeamShared's server makes the model call and every tool call for it. The sandboxed process is given a path, a home directory, a temporary directory, and the run id, and each run has its own directory. It is capped on memory, CPU time, and how many processes it may start. When the worker itself runs as root, the sandboxed process runs as an unprivileged account.
The tools a run may call are an allowlist. An agent's tool list can only remove tools. It cannot add a tool the deployment does not already allow. By default a run pauses for a person before it publishes a file, deletes, archives, or unpublishes memory or files, or starts another agent run. Those pauses stay on unless the organization names an override for that one pause. These pauses apply to agents, not to people working in the console.
Secrets
Secrets live in a vault: one for a person, one for an agent, and one for the organization. A value is encrypted before it is stored. Listing or describing a secret returns its name and a fingerprint, never the value. A grant names the tools and hosts that may use an entry, and a grant can only narrow what the entry already allows. The agent's sandboxed process does not hold the value. TeamShared's server can send it on an HTTP call, replace a placeholder the run is holding, or put it in the environment of one shell command when that entry allows it. An entry that mints a short-lived credential never gives out the stored secret. Under the default automation setting, the first use of a vault secret does not wait for a person. An organization can require that review.
Recalled memory
Shared memory is visible to the agents in the organization. Each recalled item says who wrote it: a person, an agent working with a person, an unattended agent, or the system. When a model is shown a write from an unattended agent, that write is quoted and labeled as data, not given as an instruction, unless a person has approved it.
When text is shown to a model, characters that do not display (hidden tag characters, bidirectional overrides, and zero-width marks) are written out as visible codes. Text that copies the markers the server puts around a quote is rewritten so it cannot close that quote or open a section the server wrote. The stored memory is unchanged. Ordinary text, emoji, and right-to-left writing are left as written.
Export and deletion
Org owners and admins can download memory from Export. The JSON download is the organization's active memories and playbooks, including rows that are private to a member. The markdown zip holds only memories every member can read, masks secrets in the text, and writes hidden characters out as visible codes. The vault is not in either download. An export is limited to six an hour and to 50,000 active items. An admin can mark one member's memory deleted from Organization settings. That step does not remove the rows, you cannot erase your own records that way, and the form asks you to confirm. It is limited to twenty an hour. To close an organization, sign in and purge, or email the address you used to sign up and ask us to delete it. The privacy page says the same.
On your own machines
TeamShared can run in your own infrastructure, as a containerized install. The On-prem Enterprise Deployment on the pricing page is custom and arranged with us. It is not a plan you check out, and it has no published price.
Report a vulnerability
Email [email protected]. Describe what you found and how to reproduce it. Do not include a secret, a session cookie, or an API key in the message.