mcp-knowledge-base

A Model Context Protocol (MCP) server that gives an AI assistant a searchable, persistent knowledge base. Documents are stored in SQLite and searched with FTS5 full-text search ranked by relevance.

It is a self-contained MCP server: it owns its own data and logic rather than wrapping an external API. The SQLite database is the application; MCP is the interface an AI host (Kiro, Claude Desktop, any MCP client) talks to it through.

Features

• Tools — add, search, get, list, update, and delete documents.
• Resources — read documents by URI (kb://documents, kb://documents/{id}).
• Prompts — reusable templates (summarize_search, find_related).
• Full-text search — SQLite FTS5 with BM25 relevance ranking, not naive LIKE.
• Input validation — Pydantic field constraints surfaced in the generated tool schemas.

Installation

Once published, add the server to your MCP host's config. No clone, no paths:

{
  "mcpServers": {
    "knowledge-base": {
      "command": "uvx",
      "args": ["sqlite-kb-mcp"]
    }
  }
}

• Kiro — put this in .kiro/settings/mcp.json (workspace) or ~/.kiro/settings/mcp.json (user).
• Claude Desktop — put this in claude_desktop_config.json.

uvx downloads and runs the published package on demand, so users never touch the source.

autoApprove is intentionally omitted here: auto-approving tool calls is a personal trust decision. Add it in your own config if you want read-only tools to run without a prompt.

What it exposes

Tools

Resources

Prompts

Architecture

MCP host  ──stdio──►  server.py  ──►  db.py  ──►  knowledge.db
(Kiro/Claude)         (MCP layer)     (logic)     (SQLite + FTS5)

• server.py — the MCP interface. Each tool/resource/prompt is a thin wrapper that validates input and delegates to db.py. No business logic lives here.
• db.py — pure data-access logic. Knows nothing about MCP, so it is independently testable. Opens a connection per call to avoid SQLite threading issues.
• SQLite + FTS5 — a documents table plus a documents_fts external-content virtual table. Triggers keep the search index in sync automatically, so CRUD touches only the main table.

This separation of concerns means the whole data layer is tested without any MCP machinery.

Development

# clone, then:
uv sync                 # install deps (incl. dev group)
uv run pytest           # run the test suite
uv run mcp dev src/mcp_knowledge_base/server.py   # open the MCP Inspector

To run the server directly over stdio:

uv run sqlite-kb-mcp

Notes and future work

• Transport — currently stdio (local subprocess). Streamable HTTP would make it deployable as a remote service; FastMCP supports it with a small change.
• Search — FTS5 matches whole tokens (index != indexes). Prefix queries (index*) work today; a Porter stemmer could be enabled for automatic stem matching.
• Not implemented (by design) — client-side primitives (sampling, elicitation, roots) and OAuth authorization, which are relevant mainly to remote/HTTP deployments.

License

MIT