You watch the terminal scroll past in a blur of green text. The agent is supposed to be writing a simple JSON parser. Instead, it is reaching out to an external server you never authorized. You kill the process, but the damage is done. You spend the next hour auditing every line of code to see what else it touched.
Coding agents rely on a cycle. They write, run, read the error, and fix. But most languages assume the agent is a human developer with a conscience. They allow any code to reach for the network, the file system, or your environment variables.
Vow is a programming language built for AI agents to write. It treats the agent like the machine it is.
Every function in Vow declares what it needs. A function can specify it uses net, uses fs, or uses env. If the code is not explicit about its capabilities, the language does not let it run.
When you use the command vow --sandbox, the runtime inspects the entire program before a single line executes. If the agent tries to reach beyond what you allowed, the runtime stops it cold. It refuses the program before it runs.
$ vow --sandbox --allow fs=./data app.vow
x SandboxError: needs capabilities
the sandbox does not grant
env: fn leak declares 'uses env'
Nothing was run.
This is not a suggestion. It is a hard lock. On November 12, 2026, a sandbox test of seven attack programs showed Vow blocked seven out of seven. Vow even includes a read-only file grant, letting you specify exactly which directories are safe to touch, and it blocks any code that attempts to write a file.
Vow works with the loop coding agents run. It has no nulls and no exceptions. Functions return Result values that must be handled. If an agent tries to ignore a failure, the program refuses to build.
fn parsePrice(text) {
let n = num(text)?
if n < 0 { return err("can't be negative") }
return ok(n)
}
The language is small, sitting at about 6,400 lines of JavaScript. It is open source under the MIT license. You can install it with npm install -g github:JakeBurwick/vow and find the source code at github.com/JakeBurwick/vow.
If you use Claude Code, you can add an MCP server with claude mcp add vow -- vow-mcp to give your agent the tools to guide, check, and run Vow code. You can find the documentation, a playground, and an llms.txt file to help your agents at https://vow-lang.vercel.app.
The next time your agent tries to leak your environment variables, you will know exactly what it tried to do because the sandbox stopped it at the front door.