DEV Community

Jake Burwick
Jake Burwick

Posted on

POV: Your coding agent just asked to touch your environment variables

You are staring at a terminal window. A few seconds ago, you asked your agent to parse a CSV file. Now, the agent has written a function named sync_data that clearly wants to pull your private API keys from the local environment. You are one command away from a security headache.

Coding agents operate in a repetitive loop. They write, they run, they read the error, they fix. Most languages allow any code to reach out for the network, the file system, or your system environment. If an agent decides it needs to reach into those spaces, the language lets it happen.

Vow is a programming language built for AI agents to write. I built it to keep that agent loop predictable and safe. The source code is public at github.com/JakeBurwick/vow and the project is open source under the MIT license.

Every function in Vow must declare its requirements. If a function needs to use the network, files, or the environment, it uses net, uses fs, or uses env. If the code does not explicitly declare a capability, it cannot access those resources.

The command vow --sandbox checks these declarations before anything runs. If an agent tries to reach beyond the bounds you set, the sandbox refuses the program before it starts.

$ 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.
Enter fullscreen mode Exit fullscreen mode

In a sandbox test rerun on November 12, 2026, Vow blocked seven out of seven attack programs. It also includes a read-only file grant. Using --allow fs-read=dir allows a program to read files in that directory, but if a program tries to write a file, the sandbox refuses it immediately.

Vow also addresses the agent loop by replacing nulls and exceptions with result values. Functions return results that must be handled with match or the ? operator. If an agent skips a failure check, the program will not build.

Errors are designed for agents. Each one explains what is wrong, where it happened, and the likely fix.

x TypeError: can't add a num and a str
-> line 6, col 10
6 | return "Total: " + sum
           ^^^^^^^^^^^^^^^
Likely fix: convert the number to text
first: str(sum)
Enter fullscreen mode Exit fullscreen mode

Vow compiles to JavaScript. You can install it globally with npm:

npm install -g github:JakeBurwick/vow
Enter fullscreen mode Exit fullscreen mode

If you use Claude Code, you can add an MCP server to give your agent specific tools:

claude mcp add vow -- vow-mcp
Enter fullscreen mode Exit fullscreen mode

This adds vow_guide, vow_check, and vow_run to your agent. You can find the guide, a tutorial, and a playground at https://vow-lang.vercel.app. There is also an llms.txt file and an AGENTS.md you can drop into your project so any coding agent knows how to work with Vow.

The next time your agent suggests a function that needs your environment variables, you do not have to guess if it is safe. If you do not grant the permission, the code simply will not run.

Top comments (0)