Skip to content
MCP Server

Your catalog, wired into AI assistants

DeltaForge ships a built-in MCP server (Model Context Protocol). Plug it into Claude, Cursor, or GitHub Copilot Chat once and those tools can browse your catalog, trace lineage, validate SQL, and inspect pipelines, all under your existing roles and audit log.

Catalog discovery, SQL validation, lineage, pipeline inspection
Scoped tokens, RBAC enforcement, tool-call audit logging
Runs inside your network: the AI sees results, not your storage credentials
Claude Cursor Copilot Your Agent MCP Gateway RBAC+Audit DeltaForge Catalog SQL Engine Lineage Pipelines Audit Log AUDIT TRAIL [09:14:02] claude.tool ALLOW extract_lineage(orders_fact) [09:14:09] copilot.tool DENY drop_table(prod.customers) [scope]

What the tools give the assistant

Not a thin wrapper around execute_sql. Purpose-built tools so the assistant uses typed metadata instead of inventing SQL against tables it has never seen.

Catalog discovery

The assistant can browse zones, schemas, tables, views, columns, row counts, and version history. It writes SQL against your real catalog, not hallucinated column names.

SQL validation and planning

The engine runs the actual DeltaForge parser and returns exact line and column positions for every error. Query plan inspection with row estimates lets the assistant fix plans before you run them.

Lineage and impact analysis

The assistant can trace which pipelines and views depend on a column before suggesting a refactor. Upstream and downstream traversal in one call. No more "I renamed it and the report broke."

Pipeline inspection

Read pipeline source SQL, check run status, review schedules, and get a workspace overview. The assistant can triage a failed run, read the source, validate the failing statement, and propose a fix.

Trust model

The MCP server is a thin adapter over the same control plane your humans use. The same RBAC, the same audit log, the same scoped-token model.

RBAC on tool calls

If a role cannot SELECT a column, the model cannot either. No second permission system to configure or keep in sync with the main one.

Scoped tokens

Issue per-agent tokens with explicit tool allow-lists, schema scopes, and expiration. A read-only agent literally cannot reach DROP, INSERT, or pipeline mutation.

Tool-call audit logging

Tool name, arguments, identity, and outcome land in the same audit log as human queries. Reconstruct exactly what an agent saw and did.

Runs in your network

The MCP server runs inside your DeltaForge deployment. The model provider receives prompts and tool results, not your storage credentials or raw table data.

Wire it up in one command

Install delta-forge-mcp, then run one install command per tool. It writes the config to Claude Desktop, Cursor, or Codex automatically.

# Install
brew install delta-forge-mcp

# Wire into your AI tools
delta-forge-mcp install claude
delta-forge-mcp install cursor
delta-forge-mcp install codex

# Claude Desktop config written automatically:
# { "mcpServers": { "deltaforge": { "command": "delta-forge-mcp" } } }

Give your AI assistant a job, not a sandbox

Install delta-forge-mcp and Claude, Cursor, or Copilot can work against your actual catalog under your existing access controls.