Read-only MCP server that turns a folder or repository into redacted, token-efficient context.
Model Context Protocol (MCP) Server: io.github.Avazbek22/devprojex
The io.github.Avazbek22/devprojex MCP server is a read-only service that converts a folder or repository into redacted, token-efficient context. Its purpose is to prepare context derived from existing code locations while removing or concealing selected information to reduce token usage.
🛠️ Key Features
Read-only operation
Transforms a folder or repository into context
Produces redacted output
Generates token-efficient context
🚀 Use Cases
Supplying context to MCP-based tools from an existing code folder
Creating smaller, redacted context from a repository for downstream processing
⚡ Developer Benefits
Reduces token consumption by generating token-efficient context
Ensures context is redacted for information minimization
⚠️ Limitations
Read-only: does not modify the source folder or repository
🏆 Selected by the Avalonia UI team for the App Showcase
Clean, token-efficient codebase context for AI chats and coding agents. You see exactly what leaves your project, and secrets are masked before it goes.
DevProjex turns any folder or Git repository into clean context for AI. Paste it into ChatGPT or Claude, or let Claude Code, Codex, and Cursor read the project through its read-only MCP server. A desktop app, a Terminal Workspace, a CLI, and the MCP server share one engine, one set of filters, and one masking pipeline, on Windows, macOS, and Linux.
Tick what you need in a file tree, check the live preview and its token estimate, then copy or export the result as ASCII, Markdown, JSON, or XML, or as a clean copy of the project in a folder or ZIP.
🔒 Read-only and telemetry-free by design. DevProjex does not upload your project contents or collect telemetry.
App demo 🖼️
DevProjex demo: tick files in the desktop app, preview the exact context with its token estimate, mask secrets, compress code, then do the same in the Terminal Workspace
The CLI, the Terminal Workspace, and the MCP server run straight from npm or NuGet, or from a direct release binary:
shell
npx devprojex tree . # Node.js 20+
dnx devprojex tree . # .NET SDK 10
./DevProjex tree . # release binary (DevProjex.exe on Windows)
The npm and NuGet packages contain the CLI, the Terminal Workspace, and the MCP server, but not the desktop app. Requirements per platform, the direct headless archives, and the non-root Docker image are all covered in Docs/Installation.md.
Why DevProjex? 💡
Copying files into a chat window one by one doesn't scale. Packing tools are fast, but you see the result after it's packed, not while you choose. And a coding agent with plain file access reads .env as easily as Program.cs, unless every tool it uses is told to ignore it.
DevProjex works differently: visual, precise, and local-first.
You see what leaves your project: file tree, live preview, token estimate, and an agent journal that keeps a receipt of every MCP session.
You control what leaves it: Smart Ignore, Git modes, secret and private-data masking, name filters, saved profiles.
Your agent gets the same rules: the MCP server is read-only, secret masking there has no off switch, and with Live Context the files you tick are the agent's focus.
You get more than text: a real project copy as a folder or ZIP.
The numbers, with their caveats. In the final blind series against Repomix (11 tasks, 8 repositories, 6 languages, 22 sessions per side, Claude Haiku 4.5 as the agent and as the judge), DevProjex used 973,020 tokens against 1,174,907, and the judge trusted the DevProjex answer in 17 pairs of 22 against 3. With seven planted credentials, neither tool leaked a value, but the code around them stayed readable through DevProjex in 7 files of 7 against 1 of 7. The tasks were written by the author, Repomix needs fewer tool calls, and it wins broad "how does X work" questions. The whole series, with the agents' and the judge's verdicts verbatim, is in Docs/Benchmark-History.md.
Use it for
AI assistants: clean input for ChatGPT, Claude, DeepSeek, Qwen, Kimi.
AI agents: any MCP client reads your project through a server that only reads and always masks, and you steer its focus with checkboxes.
CI pipelines: fail a build when secrets would leak into packed context, without ever printing the values.
Restricted environments: get the same clean context where AI agents, remote indexing, or IDE plugins are not allowed.
Large codebases: pull out one module instead of the whole repository.
Code reviews and teaching: share the structure and only the files that matter.
Get started
Open or drop a project folder.
Choose folders, files, filters, ignore rules, and output mode.
Preview, then copy, export, run the same workflow from the terminal, or connect your coding agent through MCP → Live context.
Works with any language, repository, or project structure.
Feature overview ✨
Choose and control
Smart Ignore: hides build output, dependency folders, and caches, only when it finds evidence.
Hide Secrets: masks detected credentials in place, in every export and every MCP response.
Hide private data: optionally masks emails, IPs, MACs, phones, and user paths.
Code compression: keeps signatures, empties function bodies; pure code gets about 3× smaller.
Desktop app, Terminal Workspace, CLI, and MCP server: the same engine, four ways to work.
Git tools built in: clone by URL, switch branches, update cached copies.
Local profiles remember your settings per project.
Light, dark, and system themes; platform-native keyboard shortcuts; localization in 20 languages.
Stays smooth even on very large folders.
Project copy export
Use File → Export project → To folder… or To ZIP archive… to create a separate copy of your current selection. The same workflow is available through devprojex export project.
Project copies respect your chosen root folders, file types, ignore rules, and checked items; if nothing is checked, the whole current tree is exported. Directory structure, binary files, and included empty folders are preserved. When Hide Secrets or Hide private data is enabled, detected values in text files are replaced, binary files remain unchanged, and the result is intentionally not a byte-for-byte copy, so it may not build or run. The source project is never modified, and the result can't be written inside it.
MCP server 🤖
DevProjex ships a built-in secure Model Context Protocol server, on the official C# MCP SDK, that turns any local folder or Git repository into token-efficient codebase context for AI agents: Claude Code, Codex, Claude Desktop, Cursor, Google Antigravity, VS Code, or any MCP client. It runs on the same engine as the desktop app, the Terminal Workspace, and the CLI. One line connects Claude Code, with or without the desktop app installed:
Under the hood every client starts the same command, devprojex mcp --root /path/to/project, and talks to it over stdio.
Connect a client
Desktop app:MCP → Live context or MCP → Standard, then Open in Claude Code, Codex, Cursor, or VS Code. The submenu you pick decides whether the registration follows the checked tree.
Terminal Workspace:mcp connect <client> [live|standard] registers the server without opening another terminal; live mode is the default here.
CLI:devprojex mcp connect . --client <client> (standard mode unless --mode live is given). Add --open to launch the client, or --print when only a manual configuration fragment is needed.
Without the desktop app: the npm or NuGet package runs the same server through npx -y devprojex or dnx devprojex. Snippets for every client are in Docs/McpServer.md.
Cursor and VS Code get a project-local .cursor/mcp.json or .vscode/mcp.json; other servers in those files are left alone. For a first-call check, inspect DevProjex with /mcp in Claude Code or run codex mcp list, then ask: “Through DevProjex, show the tree of the current selection.” That call appears in the agent journal.
Live Context: your checkboxes become the agent's focus
Open the project in the desktop app and connect through MCP → Live context. From then on the tree you tick is what the agent works with:
Tick, untick, done. The agent sees the new selection on its very next call, through the shared local project profile. No restart, no new prompt.
Focus, not a cage. Tree, search, analysis, dependency, and pack operations stay inside the checked selection. A file outside the ticks can still be read by name, with an outside-focus notice.
Your filters stay the limit. The desktop app's filters and secret masking apply to everything the agent reads; files your filters exclude stay out of reach.
Saved results stay honest. A stored pack is pinned to the selection that created it; change the ticks and the agent rebuilds it explicitly.
See what the agent took. Every file the agent received is marked in the tree with a count, the window title names the active client or session count, and focused paths are restored when you reopen the project.
Live context uses --live; Standard connects the same server but does not follow the window selection.
Eight read-only tools
Eight read-only tools cover the whole workflow: list_projects, get_tree, analyze, search_project, related_files, get_file, pack_context, and read_pack.
search_project finds text or declarations by name, and get_file reads an exact line range or one declaration instead of a whole file.
related_files answers "what does this file actually use, and who uses it" from real imports, for one file or a handful of seeds, so following a thread never widens the selection.
pack_context packs a selection under max_tokens: files are admitted in a deterministic order while they fit, the rest are reported as skipped, and top_files shows where the tokens go.
pack_context and related_files store an oversized result as a session pack; read_pack reads it back in line ranges instead of flooding the agent's context. Long operations report standard MCP progress notifications.
Trees default to compact Markdown, content declares the root once and uses relative paths, and every tree ends with a line naming the active filters, so the agent knows what it did not see. git_scope narrows any tool to staged files, current changes, or a ref-to-ref diff; profile switches between built-in defaults, your saved desktop selections, or a portable profile file.
Agent journal: a receipt for every session
The local agent journal records metadata and counts for MCP sessions, never file bodies or detected secret values. Each session shows the client, the mode, and its totals: calls, characters, token estimate, delivered files, masked values, and duration. Open a session and every tool call is listed with its arguments summary, revision, timing, and notices. Read it from the desktop app (MCP → Journal), the Terminal Workspace, or the CLI, and export a Markdown or JSON context receipt showing calls, delivered paths, and totals. Want to audit what the agent gets? Open the same project in the desktop app: the engine and the masking pipeline are shared, and filters match when the same profile and parameters are used.
Security boundaries
The server enforces hard security boundaries on top of DevProjex's read-only design:
Read-only by design. Tools cannot modify project files or run project code; network access is disabled unless --allow-remote is explicitly enabled for Git URL projects.
Secret masking is always on in MCP mode and has no off switch. Not in the server flags, not in the tool schemas, so neither a config mistake nor the agent itself can turn it off.
Optional private-data masking via devprojex mcp --hide-private-data, mirroring the CLI flag.
Root jail. Local access is pinned to startup roots; opt-in remote Git URL checkouts are pinned on first use; symlink and junction escapes are rejected.
The agent can only narrow the view. Agent paths and globs can only narrow the selection, and the .git administrative area is never exposed.
Git's view by default. Smart Ignore and .gitignore apply, while Dockerfile, .github/, dot-files, and empty files stay visible. Widening that view is your startup decision, never the agent's: an --exclude baseline, the --unrestricted preset, or the opt-in --allow-agent-exclusions flag.
Prompt-injection hardening. Returned file contents are wrapped in untrusted-data markers.
The missing off switch is a control guarantee, not a detection guarantee. DevProjex detects common secret formats, but detection is heuristic; review each pack before publishing it outside your environment.
See Docs/McpServer.md for client setup, the full tool reference, and the security model. The measurements behind the MCP server, from 5.1 to 5.2, with the agents' and the blind judge's verdicts verbatim, are in Docs/Benchmark-History.md.
Command line ⚙️
DevProjex also runs from the terminal: a keyboard-first Terminal Workspace for interactive work, and a CLI for repeatable, script-friendly project analysis and AI-context export.
bash
devprojex # open the Terminal Workspace
devprojex open . --preview # desktop app on this folder, preview pane open
devprojex tree . # project tree
devprojex tree https://github.com/owner/repo --branch main # tree of a remote repository
devprojex analyze . --format json # files, lines, tokens, largest files
devprojex analyze . --git-mode tracked --exclude smart-ignore # tracked files only, with Smart Ignore applied
devprojex analyze . --hide-secrets --findings --fail-on-findings # CI gate: findings, never values
devprojex export context . --format markdown -o ../context.md # packed context for a chat
devprojex export context https://github.com/owner/repo -o - # straight from a Git URL to stdout
git diff --name-only | devprojex export context . --select-from - -o - # only the changed files
devprojex export project . --as zip --hide-secrets -o ../submission.zip # masked project copy
devprojex search Configure . --symbols # declarations by name
devprojex related src/App.cs --direction both # what a file uses, and who uses it
devprojex profile save . --hide-secrets on # remember these settings for this project
devprojex cache update https://github.com/owner/repo # refresh a cached clone
What the CLI adds
Scriptable exports for CI pipelines: JSON reports, stdout output, deterministic files.
Git repository URLs as project sources: analyze, export, open, or start the Terminal Workspace on a repository straight from its URL through a managed clone cache (devprojex cache, devprojex recent).
A secret pre-flight for CI:--findings lists rule, category, file, and line, never the values, and --fail-on-findings fails the pipeline when findings exist.
Composable selection: pipe a file list from git diff or any tool into --select-from -.
Project search and relations:devprojex search finds text, --regex patterns, or declarations by name with --symbols; devprojex related lists the files a file uses and the files that use it.
Documented aliases and short flags (export ctx, export proj, -f, -n, -q), plus devprojex help <command> and shell completion for bash, zsh, fish, and PowerShell.
A keyboard-first Terminal Workspace with a searchable Action Palette, no desktop app needed, and the same Git modes, exclusions, and profiles as the desktop app.
DevProjex only reads your project. The one exception is an explicit Cursor or VS Code MCP connection, which creates or updates .cursor/mcp.json or .vscode/mcp.json. Beyond that, DevProjex:
Does not edit, rename, move, or delete files in the opened source project.
Does not commit, merge, push, or switch branches in the source repository opened from the user's filesystem; branch operations are limited to application-owned cached clones.
Does not include binary file contents in text or AI-context output; masking rules do not scan binary data.
Writes generated files and project copies only to destinations you choose, outside the source project.
How Smart Ignore works 🧠
Smart Ignore is a local, deterministic filter: not an AI model, and not one big blacklist for the whole folder. It knows where each project starts and ends, and applies the right rules only inside that project.
It knows project boundaries. DevProjex finds project markers like .csproj, package.json, pyproject.toml, go.mod, or Cargo.toml. Rules for that stack (node_modules, bin/obj, virtual environments, build caches) apply only inside the project that owns them. A .NET service won't hide bin in an unrelated folder next to it, and a frontend app won't hide node_modules somewhere it doesn't belong.
It checks before it hides anything. Folder names like build, dist, or vendor can mean generated files or real source code. Smart Ignore looks for real signs first: package files, compiler output, known build layouts. If there's no clear sign, the folder stays visible.
It keeps monorepos separate. Each nested project (frontend, backend, tools, docs) is filtered on its own, based on its own markers. Nothing crosses over between unrelated parts of your workspace.
Git modes are a separate setting, not part of Smart Ignore:
Mode
What it shows
.gitignore mode
Tracked files plus untracked files that repository-local .gitignore and info/exclude allow, following git status; embedded repositories stay opaque and declared submodules keep their own rules
Tracked-files-only
Only files currently recorded in the Git index
Staged
Files with staged changes
Changes
Staged, unstaged, and non-ignored untracked files
Diff
Files changed between two Git references
You stay in control. Smart Ignore, the Git mode, and basic filters (hidden files, dot-files, empty folders) all work together and can be turned on or off one by one. "Ignored" means excluded from the current view, copy, or export; it never deletes anything from your project.
📖 Full technical details, including signature matching, worktree handling, and edge cases, are in Docs/SmartIgnore.md.
How code compression works 🗜️
An AI rarely needs every implementation line; it needs the shape of your code. Compression keeps declarations, signatures, fields, and class structure, and empties named implementation bodies:
csharp
// Before: what's in your file (unchanged on disk)private Command BuildMcpCommand()
{
var command = new Command("mcp", L("Terminal.Command.Mcp"));
// … ~40 more lines of implementation
}
// After: what goes into the packed contextprivate Command BuildMcpCommand()
{ }
Measured on DevProjex's own C# sources (619 files), the packed context shrinks by about 69%, roughly 3× smaller. On a mixed repository the saving is lower, because compression only touches code, never test fixtures, JSON assets, or documentation.
It covers 14 languages, parsing with Tree-sitter grammars and applying DevProjex's own per-language rules about what must survive (properties in Kotlin, val/var in Scala, free lambdas everywhere), and it is deliberately conservative: a file it can't process safely stays complete. Comment and blank-line stripping extend the same syntax engine to 20 packs in total, the 14 plus six comments-only packs such as HTML, CSS, YAML, and XML project files. The same transformed content feeds token metrics, context documents, and folder or ZIP exports, in the desktop app and via --compress-code in the CLI.
Two independent layers. Both replace values inside the output; they never delete a file and never modify your project.
Hide Secrets finds credential values (API keys, tokens, connection strings) and masks each one where it stands, keeping the file and all surrounding code:
text
var botToken = "110201543:AAHdqTcvE…"; // in your file
var botToken = "DEVPROJEX_REDACTED[telegram-bot-api-token#1]"; // in the export
Detection runs a pinned, reviewed Gitleaks rule set on DevProjex's own engine, and one secret never costs you the whole file. Findings are marked on the preview scrollbar, and a false positive can be excluded per match right in the preview. You enable it with one switch in the desktop app or --hide-secrets in the CLI; the MCP server keeps it on, see Security boundaries. For pipelines, --findings prints rule, category, file, and line, never the value, and --fail-on-findings fails the build.
Hide private data is the second, optional layer for personal traces: emails, global IPs, local-user paths, MAC addresses, and international phone numbers become DEVPROJEX_REDACTED[email#1], [ipv4#1], [local-user#1], and so on, with the same placeholder for the same finding on every surface, so exported code stays consistent. It is off by default: real projects are full of version strings that look like IPs and sample emails that are not personal. Turn it on per profile in the desktop app, or with --hide-private-data in the CLI and MCP.
Built with .NET 10 and Avalonia UI, with Tree-sitter grammars, Gitleaks rules, and the official MCP C# SDK; attributions in THIRD-PARTY-NOTICES.md. 20,000+ automated tests (unit, integration, terminal, UI) run on Windows, Linux, and macOS.
Build from source
bash
git clone https://github.com/Avazbek22/DevProjex.git
cd DevProjex
dotnet build -c Release
dotnet test
Contributing 🤝
Issues and pull requests are welcome. Good places to start: UX, performance tuning, tests, localization, documentation and screenshots. Not sure where? Check issues labeled good first issue.
DevProjex is licensed under the Apache License 2.0: free to use, modify, and distribute, with an explicit patent grant. Copyright (c) 2025–present Avazbek Olimov. See LICENSE for details.