SIAgentNews.com
← All news

New free tool EgressVeil screens what coding agents send to AI providers

A developer released a local gateway that checks coding-agent requests for secrets before they leave your machine. It is in beta, and its protection has limits.

· First reported by dev.to

The short version

  • EgressVeil is a free, MIT-licensed, local-first gateway for AI coding agents
  • It inspects outgoing requests for API keys, passwords, JWTs, tokens and personal information
  • Four policies are available: ALLOW, REDACT, BLOCK and LOG_ONLY
  • Live API validation is done for Claude Code and Aider; the tool is still in beta
  • It covers only supported LLM API traffic sent through the gateway, not every agent action

Why it matters

Coding agents read the files in your project, and anything in those files can travel to a cloud provider along with your prompt. A tool that checks requests before they leave is one more safeguard. The basics still matter most: keep credentials out of project folders and rotate keys an agent may have seen.

What was released

A developer has published EgressVeil, a free and open-source tool meant to stop sensitive data from slipping out through AI coding agents. As dev.to reported in the author's own announcement, it is aimed at agents like Claude Code and Aider. These are programs that read your code and call a cloud language model to help write or change it.

The worry is simple. An agent working in your project might send an API key (a password-like string that lets software use a service), a login credential, customer information or details about internal systems to the company that runs the model. The author built EgressVeil to reduce that risk.

How the gateway works

EgressVeil is described as a local-first privacy gateway. In plain terms, it is a small program that runs on your own computer and sits between the coding agent and the model provider's API (the connection the agent uses to talk to the cloud). Every outgoing request passes through it first, and it applies rules before anything leaves your machine.

According to the post, it looks for API keys, passwords, JWTs (signed login tokens), other tokens and personal information. You choose what happens when it finds something, from four policies:

  • ALLOW lets the request through.
  • REDACT removes the sensitive part but lets the rest go.
  • BLOCK stops the request.
  • LOG_ONLY records the event without changing anything.

It also keeps a local audit log, which the author says does not expose the raw secrets themselves. It works with Anthropic and OpenAI-compatible APIs, and runs on Windows, Linux and macOS.

What it costs and how to start

The tool is free under the MIT License, a permissive open-source license that lets people use and change the code. The author says it has no telemetry (no usage data sent back), no accounts and no subscriptions.

Installation is a Python command, pip install egressveil. You then launch your agent through it, for example with egressveil run -- claude. The code is on GitHub, and the author invites feedback, contributions and security reviews.

The limits to keep in mind

The author is open that EgressVeil is in beta. Live API validation was completed for Claude Code and Aider, so those are the combinations that have been tested against real services.

The author also says it protects supported model-API traffic that is routed through the gateway. It does not cover every network action an agent might take. If an agent does something outside that path, the gateway will not see it.

The post is a self-announcement, and it does not mention an independent audit. Treat EgressVeil as a promising lead to evaluate, not a verified fix. A tool that handles your secrets is itself something you should be able to trust, and the invitation for security reviews suggests that checking is still ahead.

Why the underlying risk is bigger than one tool

The risk exists whichever agent you use. When an agent reads your project, anything in the files it can see may be included in what it sends to the model. A config file holding a database password is the classic example.

A few habits help whatever tool you choose:

  • Keep credentials out of the folder the agent works in.
  • Use scoped keys (limited to what is needed) that expire quickly.
  • Rotate any key an agent may have seen.

A gateway like EgressVeil adds a safety net on top of those habits. It does not replace them. The author's stated aim is to make privacy a default part of AI-assisted coding rather than an afterthought, and the post ends by asking developers how they handle this today.

What to watch

  • Whether independent security reviewers examine EgressVeil and publish findings
  • Whether support extends beyond Claude Code and Aider once the beta period ends
  • How well it handles agent traffic that falls outside the gateway's covered path

Try it

Before installing anything, search your project folders for .env files, config files and notes that hold keys or passwords, and move them out of any directory a coding agent can read. If you try EgressVeil, test it first on a throwaway project.

Sources and links

First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news