Note
This MCP server provides unified documentation search for extension developers. If you are looking to manage your live Plesk server via AI, please see the official Plesk MCP Server.
State-of-the-Art (SOTA) semantic search across the entire Plesk documentation surface, optimized for sub-second latency on Apple Silicon.
Plesk documentation is spread across five separate sources: an admin guide, a REST API reference, a CLI reference, a PHP SDK, and a JS SDK. Answering a single extension development question often means searching all of them manually, cross-referencing results, and still missing the relevant section.
This server ingests all five sources, embeds them with a multilingual model, and exposes a single search_plesk_unified MCP tool. It uses hybrid search (Vector + FTS), Reciprocal Rank Fusion (RRF), and Cross-Encoder reranking to deliver high-precision results in milliseconds.
flowchart TD
Client["MCP Client\n(Claude Desktop / Cursor / etc.)"]
Client -->|"search_plesk_unified(query)"| Server
subgraph Server["FastMCP Server · Modular Architecture"]
direction TB
Main["Bootstrap · server/main.py"]
Life["Lifecycle Hooks · server/lifecycle.py"]
Tools["MCP Tools · server/mcp_app.py"]
Main --> Life --> Tools
end
subgraph Pipeline["Retrieval Pipeline"]
direction TB
E["1 · Embed query\n(Hardware-accelerated)"]
S["2 · Hybrid Search\nVector (LanceDB) + FTS (Tantivy)"]
R["3 · RRF Merge + Rerank\n(MiniLM-L4-v2)"]
N["4 · Neighbor Expansion\n(Context Enrichment)"]
A["5 · AI Synthesis\n(sampling-enabled)"]
E --> S --> R --> N --> A
end
subgraph Store["LanceDB Vector & FTS Store"]
direction LR
G["Guide"]
A_["API"]
C["CLI"]
P["PHP Stubs"]
J["JS SDK"]
end
Tools --> Pipeline
S <--> Store
Optimized for Apple Silicon (M2/M3) using MPS acceleration and memory-resident table caching.
| Profile | Embed Model | HR@5 | MRR@5 | Avg Latency | Est. RAM |
|---|---|---|---|---|---|
light |
BAAI/bge-small | 100.0% | 0.917 | 1.007 s | ~200 MB |
medium |
BAAI/bge-base | 100.0% | 0.917 | ~0.60s | ~600 MB |
full-tq |
BAAI/bge-m3 | 75.0% | 0.750 | ~0.40s | ~1300 MB |
Metrics measured on Apple M2 Pro with LanceDB connection caching enabled.
- Single-Instance Lock: PID-based lock prevents concurrent LanceDB access when multiple MCP clients or IDE sessions try to launch the server simultaneously.
- Sub-Second Hybrid Search: Combined Vector + Tantivy FTS with RAM-cached table connections for instant retrieval.
- AST-Aware Chunking: Uses
tree-sitterto respect class and method boundaries in PHP, JS, and TS documentation. - TurboQuant Acceleration: Fast 4-bit quantized search for the
full-tqprofile, delivering 10x lower latency for large models. - Neighborhood Retrieval: Automatically fetches adjacent chunks (prev/next) to provide complete context for grounding.
- Macro-Context Summaries: Injects file-level purpose summaries into every chunk using the
SummaryCache. - AI-Synthesized Answers: Generates concise answers from search results with structured inline citations
[1],[2].
This server provides tools, prompts, and resources. See docs/mcp-components.md for a full reference.
| Tool | Description |
|---|---|
search_plesk_unified |
Hybrid search with RRF and Cross-Encoder reranking. |
get_file_content |
Retrieve the full content of a specific documentation file. |
resolve_references |
Find all files referencing a specific symbol or topic. |
refresh_knowledge |
Re-fetch sources and update the index (incremental). |
trigger_index_sync |
Start a background indexing job. |
daemon_health |
Check readiness, hardware acceleration (MPS/CUDA), and latency stats. |
plesk://toc/api- Table of Contents for API documentation.plesk://toc/cli- Table of Contents for CLI reference.plesk://toc/guide- Table of Contents for Extensions Guide.plesk://toc/php-stubs- Hierarchical list of PHP classes.
Because this server is published to PyPI and listed on the MCP Registry, you don't even need to clone the repository to run it!
You can run or integrate the server in seconds.
Add the server config to your claude_desktop_config.json (typically at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows):
{
"mcpServers": {
"plesk-dev-docs": {
"command": "uvx",
"args": ["mcp-plesk-dev-docs"]
}
}
}Go to Settings > Features > MCP, click + Add New MCP Server:
- Name:
plesk-dev-docs - Type:
command - Command:
uvx mcp-plesk-dev-docs
If you want to modify the source code, run benchmarks, or manage database migrations:
Quick bootstrap (recommended):
git clone https://github.com/barateza/mcp-plesk-dev-docs.git
cd mcp-plesk-dev-docs
./install.sh # Linux / macOS
# powershell -ExecutionPolicy Bypass -File install.ps1 # WindowsManual setup:
git clone https://github.com/barateza/mcp-plesk-dev-docs.git
cd mcp-plesk-dev-docs
uv pip install -e ".[dev]"-
Run Initial Indexing: Generate the offline vector database and full-text search indexes:
uv run python -m mcp_plesk_dev_docs.server.main refresh_knowledge
-
Start the Server:
uv run python -m mcp_plesk_dev_docs.server.main
Set environment variables in .env:
PLESK_MODEL_PROFILE=light # light | medium | full-tq
PLESK_ENABLE_SAMPLING=true # AI-Synthesized answers
PLESK_DAEMON_AUTO_WARMUP=true # Preload models on startup
PLESK_INDEX_SUMMARIES=true # Enable file-level summaries
OPENROUTER_API_KEY=sk-or-v1-...- docs/benchmarks.md - Detailed latency and quality reports.
- docs/mcp-components.md - Full tool and resource reference.
- docs/turboquant.md - 4-bit quantization internals.
MIT. See LICENSE.
This is a personal project by Gilson Siqueira. It is not officially affiliated with, endorsed by, or supported by Plesk or WebPros International GmbH. Plesk is a trademark of WebPros International GmbH.
Portions of this repository were developed under contract for Plesk International GmbH ("Plesk") only if specifically identified as such. The MIT license above applies only to material the repository owner is authorized to license. Files or directories owned by Plesk, if any, are listed in NOTICE. If you need assurance about licensing for a particular file, contact Plesk or seek legal counsel before relying on the MIT License for Plesk-owned files.
Built to make Plesk extension development faster.