You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
T3 Code for organizations: onboarding, offboarding, data control and audit
#38624
We're looking at rolling out T3 Code across our organization, and the main thing holding us back is that there's no good way to manage it centrally. Each developer runs their own server, sets it up themselves, and whatever the agents do or leave behind stays on that machine. That's fine for a few people, but it gets hard to manage at 20, and impossible at a few hundred.
Managed policy helps, but it's set per machine through MDM (device management), it's the same for everyone on that machine, and it only reaches machines IT already manages.
What's missing for us
Onboarding. A new developer has to set up repos, secrets, provider logins, device hosts and device pairing on their own. We'd like to invite someone and have them ready to go.
Offboarding. When someone leaves, their T3 sessions, device keys, Connect links, scheduled tasks and secrets keep working until someone remembers to clean them up. Their worktrees, uncommitted changes, transcripts and temp files also stay on their laptop.
Audit. We can't see who ran which agent, what commands it ran, which files it read, or which model provider got the code. Commands don't record who sent them either.
Policy per group. We'd like to set allowed providers, models and permission modes per team, and change them without pushing a new file to every machine.
Shared secrets. Secret references exist, but there's no way to give a group access to a secret and take it away again.
Shared device hosts. Macs for iOS simulators and machines for Android emulators are expensive, and we'd like to share them across the org. Today every environment adds its own SSH device hosts, and each one needs its own Simlock token.
Repo access isn't part of this. Our Git host already handles that well.
How we're thinking about it
There are two parts to this.
Organizations on T3 Connect. Invites, groups and roles, with SSO and SCIM. Managed policy pushed per group, using the same schema you already have. Removing someone from the org revokes their links, sessions, device keys, scheduled tasks and secrets in one step.
Environments that run on the org's own machines. A small runner starts one normal T3 server per developer in its own container, from a template the org sets up. Code, transcripts and temp files then live on machines the org controls, and offboarding can simply stop and delete that developer's container. Developers keep using the desktop, web and mobile apps as today. The server itself stays single-user, so this doesn't need any multi-user logic in T3.
The template is also where shared device hosts come in. The containers run Linux, so iOS simulators have to come from Macs elsewhere anyway. The org sets up its device hosts or a Simlock pool once, and every developer's environment gets them, with its own Simlock token issued by the runner.
Audit is the part we haven't figured out yet. Agent activity is a lot of data (every command, file read and transcript), so exporting all of it probably isn't realistic. We'd need to work out what's worth keeping, at what level of detail, and where it should go. For now we're only starting with the basics: recording who issued each command.
Pieces
Record who issued each orchestration command. It's a small server change, any audit later depends on it, and it's already useful for peer links, MCP clients, guests and scheduled tasks.
A Bitwarden Secrets Manager provider (bws) next to the 1Password one. Bitwarden machine accounts can be scoped by project, which fits granting and revoking secrets per group.
t3 runner with a Docker spawner. This works without server changes, since each environment can link to Connect with the device-code sign-in you already have.
Runner secrets, mounted per developer in /run/secrets from the org's secret manager.
Shared device hosts in the runner template: SSH device hosts and Simlock pools set up once for every environment, with a Simlock token per developer.
One tunnel per runner host (relay side, so yours). With dozens or hundreds of environments, a paid tunnel for each one adds up.
Organizations on Connect (also yours): invites, groups, roles, SSO/SCIM, per-group policy and offboarding.
Paid tier
We think the organization part could be something you charge for. SSO, SCIM and central policy are what companies expect to pay for, usually per seat. It also runs on the relay, which you host and operate.
Everything that runs on the org's own machines (the server, the runner, the secret providers) could stay open source. A team that doesn't want to pay could still self-host and have each developer link their own environment. There's already a precedent, since a T3 Chat plan raises the Connect tunnel cap.
Questions
Does this fit how you see T3 Code, especially keeping the server single-user and doing isolation in the runner?
Would you want the attribution change, the Bitwarden provider and apps/runner upstream?
Is organizations on Connect something you'd consider offering as a paid tier? If so, we'd be happy to shape our pilot around what you'd want to learn from it.
Have you thought about audit logging at all? Given how much data agents produce, we'd be interested in how you'd approach it.
We're building pieces 1 to 3 in a fork and will try them out internally. Happy to share how it goes, or open PRs if you'd like them.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Background
We're looking at rolling out T3 Code across our organization, and the main thing holding us back is that there's no good way to manage it centrally. Each developer runs their own server, sets it up themselves, and whatever the agents do or leave behind stays on that machine. That's fine for a few people, but it gets hard to manage at 20, and impossible at a few hundred.
Managed policy helps, but it's set per machine through MDM (device management), it's the same for everyone on that machine, and it only reaches machines IT already manages.
What's missing for us
Repo access isn't part of this. Our Git host already handles that well.
How we're thinking about it
There are two parts to this.
Organizations on T3 Connect. Invites, groups and roles, with SSO and SCIM. Managed policy pushed per group, using the same schema you already have. Removing someone from the org revokes their links, sessions, device keys, scheduled tasks and secrets in one step.
Environments that run on the org's own machines. A small runner starts one normal T3 server per developer in its own container, from a template the org sets up. Code, transcripts and temp files then live on machines the org controls, and offboarding can simply stop and delete that developer's container. Developers keep using the desktop, web and mobile apps as today. The server itself stays single-user, so this doesn't need any multi-user logic in T3.
The template is also where shared device hosts come in. The containers run Linux, so iOS simulators have to come from Macs elsewhere anyway. The org sets up its device hosts or a Simlock pool once, and every developer's environment gets them, with its own Simlock token issued by the runner.
Audit is the part we haven't figured out yet. Agent activity is a lot of data (every command, file read and transcript), so exporting all of it probably isn't realistic. We'd need to work out what's worth keeping, at what level of detail, and where it should go. For now we're only starting with the basics: recording who issued each command.
Pieces
bws) next to the 1Password one. Bitwarden machine accounts can be scoped by project, which fits granting and revoking secrets per group.t3 runnerwith a Docker spawner. This works without server changes, since each environment can link to Connect with the device-code sign-in you already have./run/secretsfrom the org's secret manager.Paid tier
We think the organization part could be something you charge for. SSO, SCIM and central policy are what companies expect to pay for, usually per seat. It also runs on the relay, which you host and operate.
Everything that runs on the org's own machines (the server, the runner, the secret providers) could stay open source. A team that doesn't want to pay could still self-host and have each developer link their own environment. There's already a precedent, since a T3 Chat plan raises the Connect tunnel cap.
Questions
apps/runnerupstream?We're building pieces 1 to 3 in a fork and will try them out internally. Happy to share how it goes, or open PRs if you'd like them.
All reactions