Shelbymcp

Local knowledge-graph memory for AI tools, implemented as one Rust binary.

Advisaur MCP gives MCP-compatible agents durable memory across sessions. Thoughts live in a local SQLite database with FTS5 search, optional vectors, typed graph edges, project scope, trust fencing, and curated context briefs. The server performs no inference and makes no model API calls.

Quick start

Run the platform-native binary through npm:

npx -y advisaur-mcp --versionnpx -y advisaur-mcp

The default stdio server stores data at ~/.studiomoser-assistant-mcp/memory.db. Existing databases created by Advisaur MCP schema v18 open unchanged.

The server was called ShelbyMCP before its rename (ADR-0012). An existing ~/.shelbymcp folder moves to the new one the first time the server runs with its default database, and the old SHELBY_DB_PATH, SHELBY_TRANSPORT and SHELBY_API_KEY variables are still read, with a warning, for one release. npx shelbymcp still runs the server and says where it went.

Package-first integrations are included in the protected v0.4.0 release:

Use advisaur-mcp setup <client> only when the client package or marketplace entry is not available yet. The fallback supports claude-code, claude-desktop, cursor, codex, gemini, and antigravity; devin directs you to its organization-managed marketplace, while windsurf keeps the legacy local config flow. It safely merges the MCP entry and never appends instructions to a global rules file. See Agent Setup.

What agents can do

The server exposes 12 MCP tools:

• Capture, search, list, fetch, update, and delete thoughts.
• Link and traverse typed graph edges with manage_edges, explore_graph, and expand_neighbors.
• Build trusted, token-bounded context with get_brief and select_context.
• Audit the database with thought_stats.

Search returns summaries and IDs so clients can fetch full content only when needed. Project-aware operations resolve scope from MCP roots and fail closed when the project cannot be identified. Untrusted or external memories are returned inside explicit data-only fences.

The optional Forage and Onboard skills are packaged with supported clients. They run in the user's agent session; Advisaur MCP does not receive or store model credentials. Standalone prompt bodies remain available:

advisaur-mcp forageadvisaur-mcp onboardadvisaur-mcp migrateadvisaur-mcp protocol

HTTP and OAuth

Start Streamable HTTP on port 3100:

ADVISAUR_API_KEY="replace-with-a-long-random-secret" \  npx -y advisaur-mcp --transport http --host 127.0.0.1 --port 3100

Endpoints include /mcp, /health, and MCP discovery documents. Setting ADVISAUR_API_KEY enables bearer authentication and an OAuth authorization-code flow with protected-resource discovery, S256 PKCE, dynamic client registration, and resource-bound access and refresh tokens. Tokens do not expire or rotate; rotate ADVISAUR_API_KEY to revoke them. Without it, HTTP runs unauthenticated and the OAuth endpoints return 503; do not bind an unauthenticated server to an untrusted network.

Build from source

git clone https://github.com/Studio-Moser/Advisaur-MCP.gitcd Advisaur-MCPcargo build -p assistant-mcpcargo test --workspacenpm cinpm test

Rust owns the product. Node is used only for the small npm launcher, packaging scripts, and their tests. See Development and Architecture.

Memory-engine pull requests also run a deterministic contract and public retrieval gate. See Memory evaluation.

Reuse in an app

The workspace is layered so an app can embed it:

• assistant-memory is the SQLite memory engine and all memory-domain behavior.
• assistant-integrations is the current client catalog plus safe setup/status/uninstall helpers.
• assistant-mcp exposes the MCP service and reusable Axum router as a library as well as a binary.

The app can embed these crates in-process, share one database contract across macOS and iOS, and host the same HTTP router without spawning Node.

License

MIT. See LICENSE.