MCP Desktop Tools

MCP Desktop Tools provides a minimal Model Context Protocol (MCP) server and a local CLI (mcp-tools) for inspecting source trees across configured workspaces using ripgrep and git.

Features (C1)

Minimal MCP server registering search_text, git_graph, repo_map, scaffold, and open_recent tools, speaking JSON over stdin/stdout.
Workspace configuration via workspaces.yaml with environment overrides.
Security helpers to keep requests within declared workspace roots, including symlink escape detection.
Persistent disk cache for repo_map and in-process TTL cache for search_text, with per-request opt-out (--no-cache).
Bounded filesystem concurrency with tunable worker pool (--max-workers / MCPDT_MAX_WORKERS).
Rich metrics: elapsed_ms, git_cmd_ms, fs_walk_count, bytes_scanned, cache_hit, and optional stage profiles (--profile).
Ripgrep and git adapters with configurable timeouts (MCPDT_SUBPROC_TIMEOUT_MS) and warnings for truncated results.
Logging with configurable level through MCPDT_LOG or the CLI --log-level flag.
Scaffold and open_recent tools for generating project skeletons and listing recently modified files, with user template overrides (MCPDT_TEMPLATES_USER_DIR).
New: snapshot tool composes git history, filesystem stats, and safe environment markers into a single JSON artifact that can optionally be logged to MLflow.
New: Minimal Python client (mcp_desktop_tools.integrations.lab_client.LabClient) for invoking tools from *Lab agents.

Installation

See INSTALL.md for detailed instructions. In short:

pipx install .  # or python -m pip install -e .[dev]

External binaries required:

ripgrep (rg) ≥ 13.0 for search_text.
git ≥ 2.30 for git_graph.

Override discovery via MCPDT_RG_PATH / MCPDT_GIT_PATH if they are not on PATH.

Configuration

Workspaces are defined in workspaces.yaml. See CONFIG.md for schema details. Environment variables can override the configuration path and ripgrep binary.

CLI Usage

Search text with ripgrep:

mcp-tools --workspace demo search_text --query "main" --include "**/*.py" --before 1 --after 1 --json

Summarise a git repository:

mcp-tools --workspace demo git_graph --rel-path proj --last-commits 20 --with-files --json

Generate a repository map:

mcp-tools --workspace demo repo_map --rel-path proj --max-depth 5 --top-dirs 30 --yaml

Capture a repository snapshot and log it to MLflow:

mcp-tools --workspace demo snapshot --rel-path proj --run-name "$BUILD_TAG" --tag repo=proj --mlflow-uri "$MLFLOW_TRACKING_URI" --experiment homelab --artifact-path repo_snapshot.json --json

Scaffold a project using the built-in templates:

mcp-tools --workspace demo scaffold --target-rel demo --template-id pyproject_min --var project_name=demo --dry-run --json

List recently updated files:

mcp-tools --workspace demo open_recent --rel-path proj --count 20 --extensions .py --json

Use --yaml to emit YAML instead of tabular output. --profile prints stage timings to stderr and includes metrics.profile in structured outputs. All cache-enabled commands accept --no-cache; filesystem-heavy commands accept --max-workers to cap concurrency. Set MCPDT_SNAPSHOT_INCLUDE_ENV=1 to include safe environment markers (os, arch, python, tool versions) in snapshot outputs.

Server Usage

Start the server with:

python -m mcp_desktop_tools.server

The server accepts JSON lines on stdin with the following structure:

{"tool": "search_text", "input": {"workspace_id": "demo", "query": "main"}}

Responses follow the unified schemas documented in mcp_desktop_tools/schemas/*.json.

Detailed usage guides for the scaffold and open_recent tools, including template formats and GardenKeeper integration examples, are available in DOCS/SCAFFOLD.md, DOCS/OPEN_RECENT.md, and DOCS/TEMPLATES.md.

Logging

Set MCPDT_LOG (or --log-level for the CLI) to control verbosity. Logs include timing information recorded by the adapters and tool execution paths. Metrics such as elapsed_ms, git_cmd_ms, bytes_scanned, cache_hit, and trimming warnings are propagated in tool responses.

Testing

Run the quality suite with:

ruff check .
mypy mcp_desktop_tools
pytest -q --cov=mcp_desktop_tools

Integration tests expect both rg and git to be available. See SECURITY.md, POLICY.md, CONFIG.md, PERF.md, and CONTRIBUTING.md for additional guidance. Tool-specific schemas are documented in DOCS/TOOLS.md, and MLflow/*Lab examples live in DOCS/INTEGRATIONS.md.