| name | mcp-gateway-getting-started |
| description | Use when the user is setting up the mcp-semantic-gateway for the first time, asks how to install or initialize it, or wants to point a coding agent (Claude Desktop, Claude Code, Cursor, Codex, opencode, pi) at the gateway. Covers `init`, the data directory layout, and the MCP client connection block. |
Getting started with the MCP Semantic Gateway
The gateway is a single MCP server that fronts every other tool source you have — native MCP servers, OpenAPI specs, hand-authored skills, generated skills — and serves only the semantically relevant subset to the agent on each turn.
Install
pip install mcp-semantic-gateway
uv sync
The mcp-semantic-gateway console command is registered on install.
Initialize the data directory
mcp-semantic-gateway init
This creates ~/.mcp_semantic_gateway/ with a starter config.toml. Everything the gateway owns lives under that directory:
config.toml — sources, LLM provider, retrieval knobs
index/metadata.db — SQLite tool registry
index/embeddings/ — local hnswlib vectors
skills/<server>/... — generated and hand-authored SKILL.md packages
No data leaves the box during indexing. Embeddings are computed locally with all-MiniLM-L6-v2.
Point an agent at it
The gateway speaks stdio MCP via the proxy subcommand. Add this block to whichever MCP client config the agent uses:
{
"mcpServers": {
"mcp-semantic-gateway": {
"command": "mcp-semantic-gateway",
"args": ["proxy"]
}
}
}
Common config locations:
- Claude Desktop:
~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%\Claude\claude_desktop_config.json (Windows)
- Claude Code:
~/.claude.json or per-project .mcp.json
- Cursor:
~/.cursor/mcp.json
- Codex / opencode / pi: see each tool's MCP server registration docs
After restarting the client, the agent gains four gateway tools:
mcp_semantic_gateway_context — set retrieval context
mcp_semantic_gateway_find_prompts — search prompt library
mcp_semantic_gateway_find_skills — search skill catalog
mcp_semantic_gateway_get_skill — fetch full SKILL.md body
Upstream tools and skills surface dynamically through these four; the agent never sees the long tail unless it asks.
Verify it works
mcp-semantic-gateway --version
mcp-semantic-gateway search "list pull requests"
If search returns no results, you have not run mcp-semantic-gateway index yet, or no sources are configured. See the mcp-gateway-configure-sources skill.
Next steps
- Add sources —
mcp-gateway-configure-sources
- Generate skills from OpenAPI —
mcp-gateway-synthesize-skills
- Use the gateway from inside an agent —
mcp-gateway-discover-by-intent