DevProjex is a local-first codebase context builder designed to generate clean, safe, token-efficient project context for LLMs, AI chats, and coding agents. It provides a desktop app, terminal workspace, CLI, and read-only MCP server, all four sharing the same engine, the same filtering rules, and the same redaction pipeline, with support for Windows, macOS, and Linux.
Purpose Copying files into a chat window and pasting them one by one does not scale; packaging tools are fast, but the result cannot be seen until packaging is complete; and giving a coding agent direct file access means it will read .env just like ordinary source code. DevProjex's approach is visual, precise, and local-first: letting users clearly see what content will leave the project, and control the scope of what leaves.
Main capabilities
- Selection and control: Smart Ignore hides build output, dependency folders, and caches, and looks for evidence before hiding them; Hide Secrets redacts detected credential values in place in exported content and MCP responses; Hide private data can optionally mask emails, IPs, MACs, phone numbers, and user paths; code compression preserves declarations and signatures and empties implementation bodies, shrinking pure code by about 3 times; comment and blank line removal performs syntax-aware cleanup based on 20 language packs without modifying source files. The file tree supports checkboxes, search, and name filtering. Git mode covers .gitignore, tracked files, staged files, all current changes, or ref-to-ref diffs.
- Read-only MCP server: provides eight tools, including list_projects, get_tree, analyze, search_project, related_files, get_file, pack_context, and read_pack, and can connect to Claude Code, Codex, Claude Desktop, Cursor, Google Antigravity, VS Code, or any MCP client. Clients start the server via devprojex mcp --root /path/to/project and communicate over stdio.
- Live Context: after checking files in the desktop app, the agent can use that selection on its next call, with no restart or re-asking required. File tree, search, analysis, dependencies, and packaging operations are all limited to the checked scope.
- Agent session logs: Record MCP session metadata and counts locally, including call counts, character counts, token estimates, delivered files, masked values, and duration; file contents and detected secret values are not recorded. Context receipts can be exported as Markdown or JSON.
- Preview and export: Preview the tree, contents, or both in real time, and display line counts, character counts, and token estimates; scrollbar markers show search matches and redaction findings; export as ASCII, Markdown, JSON, or XML, to a file or the clipboard; project copies can be exported as a folder or ZIP.
- Command line: The terminal workspace is suited to keyboard-first interactive operations, while the CLI is suited to repeatable, scriptable project analysis and AI context export, with support for JSON reports, stdout output, deterministic files, Git repository URLs as project sources, CI secret preflight checks, and more.
How to use Open or drag in a project folder, select folders, files, filters, ignore rules, and output mode, preview, then copy or export. You can also run the same workflow from the terminal, or connect a programming agent via MCP. The desktop app connects clients via MCP → Live context or MCP → Standard. Without the desktop app, the npm or NuGet package can run the same server via npx -y devprojex or dnx devprojex. On Windows, it can be installed via winget install OlimoffDev.DevProjex.
Use cases Suitable for clean input for AI assistants (ChatGPT, Claude, DeepSeek, Qwen, Kimi), AI agents reading projects via MCP, CI pipelines failing builds when secrets leak into packaged context, restricted environments where AI agents or remote indexing are not allowed, extracting only one module from a large codebase, and sharing only structure and key files during code review and teaching.
Limitations and notes In MCP mode, secret redaction is always on and cannot be turned off, but detection is heuristic and should be manually checked before release. Project copies with redaction enabled are not byte-for-byte copies, may fail to build or run, and results cannot be written into the source project. The server is read-only, and tools cannot modify project files or run project code; unless --allow-remote is explicitly enabled, network access is disabled by default. For details such as platform requirements, headless archiving, and non-root Docker images, thedocumentation contains installation guidance, which is not covered one by one in this article.
Other facts The project was selected by the Avalonia UI team for the App Showcase. Official benchmarks (11 tasks, 8 repositories, 6 languages, 22 sessions per side, with Claude Haiku 4.5 as agent and judge) show that DevProjex used 973,020 tokens, compared with Repomix's 1,174,907; in 22 pairs, the judge trusted DevProjex's answers in 17 pairs, compared with 3; none of the seven planted credentials leaked in either, but the surrounding code remained readable via DevProjex in 7 files with 7 instances, compared with 1. The tasks were written by the author, Repomix required fewer tool calls, and it won on broad "how does X work"-type questions.
