An MCP server for your data — auto‑generated from your schema

# mcp config — Claude Code, Cursor & other HTTP MCP clients
# (stdio-only clients such as Claude Desktop connect via mcp-proxy)
{
  "mcpServers": {
    "foxnose-support": {
      "type": "http",
      "url": "https://<env_key>.fxns.io/support/_mcp",
      "headers": {
        "Authorization": "Simple <public_key>:<private_key>"
      }
    }
  }
}

Connect Claude guide →  ·  Connect ChatGPT & OpenAI →

Same auth as REST

If a Flux key can't see Collection X over REST, it can't see Collection X over MCP. Writes are double opt-in: the key's role and the target Collection must both allow them — so an agent can read everything but write only into its memory Collection. One permission system, not two.

Schema, not a copy

The describe_resource tool returns the live JSON Schema. The agent validates its calls against the current shape before issuing them. Hand-written MCP servers maintain a copy that drifts.

Constant catalog size

Seven tools regardless of how many Collections or tenants you have. Hand-rolled MCP servers inflate one-tool-per-resource and overflow the model's tool budget.

Hybrid search in one tool

Vector similarity, full-text matching, and structured filters fused server-side as a single search_records tool. The model doesn't choose between three narrow tools and pick poorly.