Skip to main content
Glama
Mnē-MCP
by Gatoco

Mnē-MCP — MCP RAG Server for opencode

Mnē-MCP (from the Greek Mnēmosynē, Titaness of memory; the macron in ē, U+0113, marks the long vowel in the transliteration) is an MCP server that implements a complete RAG (Retrieval-Augmented Generation) pipeline for opencode: it indexes an Obsidian vault or document folders and answers assistant questions using your own notes as a source, citing the exact path of each fragment used.

Stack: #RAG #MCP #mcp-server #opencode #Obsidian #Qdrant #Ollama #DeepSeek #bge-m3 #Python

Features

  1. Incremental indexing — scans directories (.md, .txt) from disk with change detection via mtime + md5, stale cleanup, and support for large volumes (~19 GB).

  2. Local embeddings — bge-m3 model via local Ollama (/api/embed), 1024 dimensions, multilingual (including Spanish), L2-normalized to scale correctly with cosine similarity.

  3. Semantic retrieval — Qdrant (Docker), single collection with payload filters (source, path, mtime, md5, heading_path) and configurable relevance threshold.

  4. Grounded generation — answers with deepseek-v4-flash (Ollama Cloud) only when evidence exceeds the threshold; otherwise returns explicit rejection ("no relevant documents found"). Without evidence, it does not generate.

Related MCP server: Obsidian MCP Assistant

Architecture

MCP tools (stdio) ──┐
                    ├── RagService (core) ────▶ Qdrant (vectorial, Docker)
Admin page (:8310) ─┘       │
                            ├── OllamaEmbedder (bge-m3 local)
                            └── LLMProvider (deepseek-v4-flash, Ollama Cloud)
  • Python 3.14 · official MCP SDK (mcp==1.12.4) · qdrant-client · httpx

  • 7 tools: index · query · search · delete · list · stats · config

  • Local admin page (http://127.0.0.1:8310): dashboard, indexed documents, asynchronous indexing with progress, and search playground. Shares the same core as the MCP tools (zero duplicated logic).

  • Suite of 70 offline tests (Qdrant :memory: + mocks) + end-to-end smoke test.

Developer Usage

docker compose up -d qdrant   # Qdrant
ollama pull bge-m3            # embeddings locales
cp .env.example .env          # OLLAMA_API_KEY + VAULT_ROOT
make install                  # instala el paquete (venv)
make admin                    # panel: http://127.0.0.1:8310

Register in opencode.json (mcp section, local type, command python3 mcp_rag/server.py or the mcp-rag-opencode entry point).

Request

Behavior

index (source, path, force_rescan, max_files)

Incremental scan, skips unchanged files, uploads vectors in batches

query (question, top_k ≤ 8, source, path_prefix, score_threshold)

Retrieval → reranker/LLM → answer with citations (path + heading_path)

search

Raw hits with score, no generation

stats

Qdrant/Ollama health, LLM status, counts by source

Roadmap

Hybrid dense+sparse search (BM25 + RRF) · cross-encoder reranking · semantic cache · evaluation harness (recall@k, MRR, nDCG) — planned in openspec/specs/mnemos-differentiators/.

Academic

Integrated Project · Prof. Christian Pérez · Student: Gat · 2026

Available Tools

7 tools
configC

Get the full config, or set one persisted field.

Inputs: key, value, action.

deleteA

Delete indexed docs by exact path or by source.

Inputs: path, source.

indexB

Index the vault (source root or sub-path) into Qdrant.

Inputs: path, source, max_files, force_rescan.

listA

List indexed documents (paginated, filter by source/path prefix).

Inputs: limit, offset, source, path_prefix.

queryC

Retrieve context and answer the question with the LLM.

Inputs: question (required), top_k, source, path_prefix, +1 more.

statsB

Return collection stats and provider health.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 7 tools

    Disambiguation5/5

    Each tool has a distinct role: indexing, LLM-based querying, raw vector search, deletion, listing, stats, and config. Query and search are clearly separated by whether the LLM is involved.

    Naming Consistency5/5

    All tool names are lowercase, single-word commands that follow a consistent imperative/noun style. The set is uniform and predictable despite not using a verb_noun pattern.

    Tool Count5/5

    Seven tools is well-scoped for a focused memory/RAG server. Each tool covers a necessary operation without redundancy or bloat.

    Completeness5/5

    The toolset covers the full indexed-document lifecycle: add via index, retrieve via search/query, remove via delete, inspect via list and stats, and configure via config. No critical missing operation is apparent.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      A
      quality
      D
      maintenance
      Enables indexing local documents (PDF, Markdown, text, code) into a knowledge base and querying them via semantic search using local embeddings, all running privately on your machine.
      4
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI coding agents to search, read, create, update, and delete notes in a local Obsidian vault through hybrid semantic and lexical retrieval, with all embedding and vector storage running locally.
      1
      MIT