codexify connects ChatGPT web to local development workflows
codexify, by Devnoname120, is a local MCP bridge server that connects the ChatGPT web interface to a developer's local project directory. It enables web models to perform coding tasks such as refactoring, running tests, and managing version control directly on local files. The app exposes file read/write, shell execution, git operations, and Secure MCP Tunneling for private connectivity. It targets software engineers and power users who require model-assisted code work while retaining control of project data.
What tasks can you actually use it for?
The app routes web-model actions into a local workspace so agents can perform concrete development work. Supported actions listed by the developer include local file read/write/search, shell command execution with stdin/stdout handling, and Codex-style agent contracts such as apply_patch and view_image. Persistent task notes and plan tracking let agents carry out multi-step refactors and maintain progress across long-running coding sessions.
How reliable are its agent-driven code changes?
The developer provides built-in git operations and project-scoped diff checkpointing so generated edits can be recorded and inspected. That structure improves traceability when models modify files or create commits, and the app can run tests as part of agent workflows. Despite these controls, generated edits reflect model outputs and should be treated as proposals that need human verification before merging into main branches.
What environment and inputs does it require?
The app is implemented in Rust and runs on Windows, macOS, and Linux; it requires a local installation and MCP-compatible clients or the ChatGPT web interface connected via secure tunnels. Installation options include a curl command for macOS/Linux and a PowerShell script for Windows, followed by a guided quickstart that configures local endpoints and a dispatcher for aggregated upstream MCP servers.
Does it keep your code and connections private?
The design emphasizes local-first control and secure tunneling. The Secure MCP Tunnel uses random per-process bearer tokens and listens only on, avoiding inbound ports and public URL exposure. Aggregating multiple upstream servers into a single dispatcher centralizes the tool surface without exposing local endpoints, supporting a workflow where project context and files remain under the user's control rather than being published externally.
Best for developers who commit to local workflows and strict review
The app is a practical option for engineers who want model-assisted edits inside their own repositories and who can validate automated changes. It demands disciplined verification, since agents can modify files and produce commits; users should make human review and automated testing routine when accepting AI-proposed edits. For teams, treat agent outputs as draft changes that require code review before integration.
