Clawvisor

@OverlordsAi When you approve a task with scopes like send or delete, you approve it either for temporary or ongoing access, but only if it is aligned with the task description. If you permit it to delete XYZ and it tries to delete ABC instead, it will be blocked automatically.

@clawvisor So I’m confused - so if I let it send/delete the XYZ, I need to approve it once and it can send/delete any time forever. How is the difference from permissions (delete-no) Humans tend to gloss over stuff like this anyway it’s like ok button in terms of conditions

Introducing Clawvisor, the authorization layer for AI agents. Before today, giving access to your agents was all-or-nothing. Now, you can give your agents access that is scoped to your specific task. If your agent drifts or is compromised, it's blocked. Try it today for free.

@aritifactual @ycombinator @ericlevine It uses OAuth, yes, but OAuth wasn't built for nondeterministic agents. In OAuth, "read email" means both "read recently received emails for triage" and "exfiltrate all emails". Clawvisor allows you to scope access to only the data your agent needs.

@ycombinator @clawvisor @ericlevine I’m assuming this is just OAuth? Or what? Can’t everyone do this already with all the other ai agents out there?

Clawvisor (@clawvisor) lets you give AI agents access to apps like Gmail and Slack without handing over your credentials or worrying they'll go rogue. You approve tasks once, Clawvisor enforce them. Congrats on the launch, @ericlevine! ycombinator.com/launches/QFP-c…

@AndrewWarner @ycombinator @ericlevine @zapier Not quite! We're big fans of Zapier, but they allow you to define more deterministic workflows. Clawvisor governs always-on agents that reason and improvise. Very different problems.

@ycombinator @clawvisor @ericlevine Isn't that what @zapier does?

Go to clawvisor.com to create a free account, or go to github.com/clawvisor/claw… to self-host.

@harikp2002 @ycombinator @ericlevine It's multi-tiered. When you connect an account, you can put in hard restrictions on what actions it can take. Then, agents need to enumerate the actions they need to complete the task. Finally, an LLM judge makes sure the usage of those actions is coherent with the task itself.

@ycombinator @clawvisor @ericlevine how are the guardrails setup? llm as a judge?