Agents act on your behalf. These are the rules they follow.
Vekrun runs commands, edits files and talks to AI providers for you. Four commitments shape how it does that.
Nothing an agent reads is trusted
Text in your repository, model replies, tool results and terminal output cannot change what an agent is allowed to do. Your permissions decide, not the prompt.
Keys stay in your keychain
API keys are stored by your operating system, hidden from logs and the interface, and never handed to tools or remote machines unless you say so.
Deny by default
An agent can only do what you have allowed. Every action that touches files, commands, Git or the network is recorded.
Limits are stated
How well an agent is isolated differs by operating system. Vekrun says what each platform can and cannot enforce instead of glossing over it.
Reporting a vulnerability
Please report security problems privately, never in a public issue.
A security address will be published here before the first public release. Once the source repository is public, GitHub's private vulnerability reporting will also work. Include steps to reproduce, your platform and the Vekrun version; reports are acknowledged within seven days.
There are no releases yet. Once there are, the latest release is supported.
What counts
- Any way for an agent, an integration or content in a repository to act outside your permissions.
- Any exposure of API keys or credentials.
- Any way for an interrupted task to repeat a command that should run once.
- Any way to reach Vekrun from outside your machine without being allowed to.
- Problems in how Vekrun is built, released or updated.