v2.7.13: Claude Employees That Stay Logged In, Webhooks You Can Actually Trust, and a Friendlier Free Tier
Every release has a personality. This one is about trust: making sure your AI employees stay logged in, making sure your integrations are actually protected, and making sure the free tier behaves the way you expect it to.
Here is what is new in v2.7.13.
Your Claude employees stay logged in, permanently
The headline fix of this release is a quiet one, and it is the kind you only notice by its absence.
A small number of AI employees running on Claude could lose their authentication after a container restart. Nothing crashed and nothing looked broken. The employee simply stopped answering, which is the most frustrating version of a problem because there is nothing to point at.
We traced it, fixed the root cause, and then went further. Employee credentials are now protected by several independent safety nets. They are placed correctly when the employee starts. They are checked again a few seconds after boot and restored automatically if anything overwrote them. And they are verified one final time every time a connection opens, which covers manual restarts and anything else outside the normal flow.
That is a lot of belt and braces for a single file, and deliberately so. An AI employee that cannot authenticate does not complain, it just goes quiet. We would rather over-engineer the thing that keeps your team online than have you wonder why someone stopped replying.
Every affected employee in production has been restored and verified, and the mechanism is validated end to end.
Webhooks now protect what they promised to protect
Webhooks are how your AI employees get triggered by the rest of your stack: a form submission, a CRM update, a deploy, a payment. If you set a secret on a webhook, you expect that secret to be the thing standing between your employee and the open internet.
Now it genuinely is. A webhook with a secret requires a valid signature, and unsigned requests are rejected. We also made the verification friendlier by accepting two formats, so you can send the raw secret in a header or a standard HMAC signature of the request body, whichever your sender supports. Both work, and both are documented with copy-and-paste examples in the webhook-manager skill.
There is a detail here we genuinely enjoy: this issue was found and reported by one of our own AI employees while working with the webhook skill. An agent audited its own integration surface, noticed the logic did not match the documentation, and filed it. That is exactly the kind of behavior we built the platform for.
Webhooks are a team feature now
Previously, only the person who created an employee could set up that employee's webhooks. If you shared an employee with a colleague, or if an admin needed to wire up an integration, they were locked out of a feature they clearly should have.
That is fixed. Owners, admins on Enterprise, and any teammate the employee is shared with can now list, create, update, rotate secrets for, and delete webhooks. Ownership stays clean behind the scenes, so a webhook created by a colleague still belongs to the employee's owner and nothing gets orphaned when people move around. The permission rule is the same one that already governs shared files, so there is one model to learn instead of two.
A free tier that gets out of your way
Three improvements here, all pointing the same direction.
Creating your first employee just works. Admins could occasionally find the create button disabled even with no employees of their own, because the count was looking at the whole visible team instead of their own. The rule is now simple and consistent for everyone: one employee per account on the free tier, counted per account, no surprises.
A quieter, snappier dashboard. The interface was generating a stream of permission errors in the background for features outside the tier. Those are gone, so everything feels lighter.
A cleaner model picker, and a menu that matches reality. Free tier users see the model name and the effort levels they can choose from, presented clearly, without provider plumbing cluttering the interface. Options that are not part of the tier are now hidden rather than shown and then refused. Fewer dead ends, fewer moments of wondering whether you did something wrong.
The thread running through all of it
The fixes in this release share a theme worth naming. Each one was a case of something that looked right rather than something that was right. A permission check that appeared to guard the door. A limit that counted the wrong thing. A file that seemed to land where it should.
Autonomous systems raise the stakes on that category of problem, because there is no one watching the screen to think "that is odd". Your AI employee is supposed to work while you are asleep, in another meeting, or on holiday. Silence has to mean nothing needs your attention, not that something quietly stopped.
That is what most of v2.7.13 is really about: making silence trustworthy again. Employees that stay logged in without being nudged. Secrets that actually gate access. Limits that count what they say they count. Menus that only offer what will work.
None of it is flashy. All of it is the difference between a platform you check on and a platform you rely on.
Ready to put an AI employee to work? Start at geta.team.
