A coding agent that never uploads your code
DevSpace inverts the usual bargain of AI coding assistants. Instead of shipping your files to a vendor's cloud, it hands ChatGPT a connection back to your own machine, so the model reads, edits, searches, and runs code in your real local projects without uploading anything to a third party. The pitch on the tin is blunt: "Turn ChatGPT into Codex."
The mechanism is a self-hosted MCP server. You install the CLI, run devspace init and devspace serve, and expose the local port (usually 7676) through a tunnel you control. The README names Cloudflare Tunnel, ngrok, Pinggy, and Tailscale Funnel as options. Your MCP client then connects to the public /mcp URL, and DevSpace gates the connection behind an Owner password that only you hold. That password is printed by devspace init and stored at ~/.devspace/auth.json.
What the model actually gets
Once connected, ChatGPT can open one of your approved project folders as a workspace. The README lists the tools DevSpace exposes inside that workspace: read, write, and edit files; search code and inspect directories; run shell commands for tests, builds, git, and package scripts; and use isolated Git worktrees for parallel coding sessions. It also reads project instructions from AGENTS.md and CLAUDE.md, and can discover local agent skills from your skill folders.
Two design choices stand out. The first is scoping: you decide which roots are allowed during setup, so the model only ever sees folders you name, not your whole disk. The second is the worktree support, which lets separate coding sessions run in parallel without stepping on each other, a detail that suggests the author expects people to run more than one agent task at a time.
The honest part about risk
The README does not soft-pedal what a remote coding agent means. Under a "Mental Model" heading it states plainly that DevSpace is remote access to selected local folders, and that the connected client still has powerful local capabilities inside an opened workspace, including shell execution. Its own advice: treat a connected client like a trusted coding partner with access to your machine. That framing is more candid than most tools in this category manage, and it matters, because "run shell commands" is exactly the capability that turns a convenience into an attack surface if the connection or the tunnel is misconfigured.
There is also a note addressing the obvious question of whether this violates OpenAI's terms. The author argues it is a standard custom connector setup and that writing or running code is not a restricted use case, while cautioning that your account is governed by your usage, not by DevSpace. Read that as the author's claim and a pointer to check your own provider's terms, not as legal cover.
Where it runs, and where it doesn't
DevSpace requires Node >=22.19 <27 and supports Linux, macOS, and Windows, with one sharp caveat: Windows needs a Bash-compatible shell. Git Bash, WSL, MSYS2, or Cygwin Bash work; plain PowerShell or cmd.exe are listed as not supported yet. A devspace doctor command exists to inspect your local setup, which is a good sign the author has hit enough environment mismatches to build a diagnostic for them.
The README carries the usual signs of a solo, opinionated project. There is a sponsor table, a plug for the author's other product, and a "Philosophy" section betting that ChatGPT becomes "the operating system for everything." None of that is load-bearing. What is load-bearing is the core idea, and it is a clean one: keep the code on your machine, keep the credentials in your hands, and let the model reach in through explicit, inspectable tools rather than a file upload.
For anyone who wants an agent-driven coding loop but cannot or will not send source to a vendor, Waishnav/devspace is a concrete answer worth evaluating. The security burden shifts onto you to tunnel and approve carefully, but that is arguably where it belongs.
GitHub: https://github.com/Waishnav/devspace
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (1)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.