mcp-digger

mcp-digger

Code context for AI coding agents. Progressive, on-demand access to your internal .NET / NuGet package source — agents browse, search, and read private C# libraries autonomously, with zero workspace pollution.

⚠ Scope: .NET / C# only. mcp-digger indexes NuGet-style repos containing .csproj packages and .cs source files. It is not a general-purpose source indexer — other languages (TypeScript, Python, Java, Go, etc.) are out of scope.

---

✨ Why

Public NuGet packages have documentation ecosystems — API references, tutorials, community Q&A. Tools like context7 serve that well.

Internal .NET packages often have source code as their primary documentation. mcp-digger turns that source into structured, searchable, token-efficient context that any MCP-compatible agent can consume — bridging the documentation gap in private C# library ecosystems.

Without it:

• 🐢 Slow context gathering — git clone + find + grep + cat chains burn tokens on infrastructure before useful context is retrieved.
• 🔍 No semantic search — file system tools find text, not API surfaces. "Every type implementing this interface" means writing extraction scripts on demand.
• 💸 Token waste — agents read whole files when a single method signature would do.
• 🖱 Permission click fatigue — dozens of shell-command approvals per session.
• 🧹 Workspace noise — referenced repos pollute file search, git status, and the agent's context window.

With it:

• ✅ Correct code on the first try — real signatures, generic constraints, interface contracts, base class patterns.
• ⚡ Self-service context — point the agent at your NuGet repos once; it browses, searches, and reads autonomously.
• 🪙 Progressive disclosure — 200-token overview before 5,000 tokens of source. Most questions resolve at L1 or L2.
• 🧼 Zero workspace pollution — managed clones live outside your project tree.
• 🌐 Any Git host — GitHub, GitLab, Azure DevOps, Bitbucket, self-hosted — HTTPS or SSH.

---

🛠 How it works

Ten purpose-built tools, escalating from broad to deep. The agent picks the cheapest tool that answers its question.

                                                           ┌─→ 📦 dig_package_overview ─┐

│ (docs, key types) │ 🩺 dig_status → 📋 dig_list → 📖 dig_repo_overview ┤ ├─→ 🔎 dig_lookup → 📄 dig_file (health) (discover) (README + summaries) │ │ (symbol → file) (full source) ├─→ 📁 dig_package_files ────┤ │ (file listing) │ └────────────────────────────┴─→ 📝 dig_signatures (stripped API)

Operational: 🔄 dig_refresh (force cache invalidation, on demand) Bootstrap: 🌱 dig_init (only when no config exists)

---

🧰 Tools (10)

Search modes for `dig_lookup`:

---

🚀 Quick start

Install

npm install -g mcp-digger

or run directly

npx mcp-digger

Requires Node.js 20+, git on PATH, and a .NET / C# source repo (NuGet packages with .csproj + .cs sources).

Minimal config

Create .digger/config.json in your workspace root:

{

"repos": [ { "name": "my-libraries", "url": "https://github.com/org/shared-libs.git", "packageFilter": "MyCompany.*", "auth": { "strategy": "pat", "PAT-EnvVarName": "GIT_PAT" } } ] }

Don't have a config yet? Start the server, then call dig_init to scaffold one.

Agent setup

Claude Code

Add to .claude/settings.json or project settings:

{

"mcpServers": { "digger": { "command": "npx", "args": ["-y", "mcp-digger"] } } }

</details>

Codex CLI

Add to ~/.codex/config.toml (or .codex/config.toml for project-scoped):

[mcp_servers.mcp-digger]

command = "npx" args = ["-y", "mcp-digger"]

</details>

Claude Desktop

Add to claude_desktop_config.json:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{

"mcpServers": { "digger": { "command": "npx", "args": ["-y", "mcp-digger"] } } }

</details>

VS Code

Add to .vscode/mcp.json (workspace) or your user mcp.json:

{

"servers": { "digger": { "type": "stdio", "command": "npx", "args": ["-y", "mcp-digger"] } } }

</details>

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{

"mcpServers": { "digger": { "command": "npx", "args": ["-y", "mcp-digger"] } } }

</details>

Verify

Once connected, ask your agent to call dig_status — it reports config validation, per-repo connectivity, and index health.

---

⚙ Configuration

Repos & packages

A repos[] entry has three ways to declare packages:

sourceRoot defaults to "src" — set it to whichever directory holds your package folders. The walk is recursive, so nested layouts like src/Group/Foo/Foo.csproj are picked up.

Branch tracking

By default, managed clones use the repo's default branch. Pin to a specific one:

{

"repos": [ { "name": "my-libraries", "url": "https://github.com/org/shared-libs.git", "branch": "develop" } ] }

The branch is used for both initial clone and subsequent fetches. Only applies to managed clones — for local repos, you control the checked-out branch yourself.

Local repos (Mode B)

Skip managed cloning when the repo is already on disk. The local path is read-only — mcp-digger never fetches or modifies it.

{

"localRepos": { "my-libraries": "C:/repos/shared-libs" }, "repos": [ { "name": "my-libraries", "sourceRoot": "src" } ] }

</details>

Auth strategies

PATs can be inline ("PAT": "...") or via environment variable indirection ("PAT-EnvVarName": "MY_TOKEN"). The .env file in your workspace root is loaded automatically — values containing # should be quoted.

Environment variables

Secrets (PAT values) belong in .env or the real environment — never as env vars in this table.

---

🩺 Diagnostics & recovery

Debug log

Enable debug logging in your config:

{ "debug": true, "repos": [...] }

Logs go to .digger/debug.log (capped at 5 MB, auto-truncated). Critical errors and crash output land in .digger/error.log.

---

💬 Feedback

Tried mcp-digger on your codebase? Share what worked, what broke, what's missing in GitHub Discussions. Bug reports go in Issues.

---

📜 License

MIT License — see LICENSE.