Convex documents a Model Context Protocol server shipped in the Convex CLI at docs.convex.dev/ai/convex-mcp-server, started locally with `npx -y convex@latest mcp start` over stdio after linking a project via `npx convex dev`. Tools include `status` (deployment selector), `tables` (declared and inferred schemas), `data` (paginated documents), `runOneoffQuery` (sandboxed read-only JS), `functionSpec`, `run` (invoke deployed functions), `logs`, `insights` (72-hour OCC and resource-limit health), and `envList`/`envGet`/`envSet`/`envRemove`. Flags such as `--prod`, `--preview-name`, `--deployment-name`, and `--disable-tools` restrict scope; production writes require `--dangerously-enable-production-deployments` per changelog defaults.
Use cases
- Let Cursor inspect Convex table schemas before writing mutations
- Paginate production-like dev data with the `data` tool during debugging
- Run sandboxed `runOneoffQuery` probes without deploying new functions
- Check `insights` for OCC conflicts after agent-driven schema changes
- Disable `run` or `envSet` via `--disable-tools` for read-only agent sessions
Key features
- Cursor
- Claude Desktop
- Claude Code
- VS Code
- Codex
Frequently Asked Questions
- Does this connect to production by default?
- Docs and CLI changelog note production access is blocked unless you pass explicit prod flags.
- Is this a separate npm package?
- No—the MCP server ships inside the official `convex` CLI (1.19.5+).
- Can agents write data?
- `run` can invoke mutating functions; use `--disable-tools` or dev-only deployments to limit blast radius.
Related
Related
3 Indexed items
ClickHouse MCP Server
The open-source ClickHouse MCP server (PyPI package `mcp-clickhouse`, repository ClickHouse/mcp-clickhouse) exposes MCP tools such as `run_query`, `list_databases`, and paginated `list_tables` against ClickHouse clusters, defaulting to read-only SQL unless `CLICKHOUSE_ALLOW_WRITE_ACCESS` is enabled. Optional chDB extras add `run_chdb_select_query` for embedded queries over files and URLs. HTTP/SSE transports require authentication via `CLICKHOUSE_MCP_AUTH_TOKEN`, FastMCP OAuth/OIDC providers, or explicit `CLICKHOUSE_MCP_AUTH_DISABLED=true` for local dev; a `/health` endpoint supports orchestrator probes without credentials per README guidance.
Milvus MCP Server
The zilliztech/mcp-server-milvus project (documented at milvus.io/docs/milvus_and_mcp.md) exposes Milvus vector-database operations to MCP clients such as Claude Desktop and Cursor. The recommended launch path is `uv run src/mcp_server_milvus/server.py --milvus-uri http://localhost:19530` without a separate install step, with optional `MILVUS_URI`, `MILVUS_TOKEN`, and `MILVUS_DB` environment variables. Tools listed in Milvus docs include `milvus-text-search`, `milvus-hybrid-search`, `milvus-multi-vector-search`, `milvus-query`, and `milvus-count` for collection management, semantic retrieval, filtered hybrid search, and entity counts.
MotherDuck MCP Server
MotherDuck documents a remote Model Context Protocol server at `https://api.motherduck.com/mcp` (fully managed, read-write) that lets Claude, ChatGPT, Cursor, Claude Code, Codex, and other MCP clients query MotherDuck cloud databases via OAuth or Bearer tokens with zero local install per motherduck.com/docs/sql-reference/mcp. Remote tools include read-only `query` and read-write `query_rw` plus schema exploration helpers documented in the MCP workflows guide. For local DuckDB files, S3 paths, or custom limits, the open-source `mcp-server-motherduck` package (motherduckdb/mcp-server-motherduck on GitHub, PyPI `mcp-server-motherduck`) runs via `uvx` with flags such as `--db-path`, `--motherduck-token`, and `--read-write` for self-hosted stdio or HTTP transports.