Coral

Coral is a read-only SQL layer for enterprise AI agents. You point it at the systems you already run — GitHub, Datadog, PagerDuty, Slack, Linear, Stripe and dozens more — and each one becomes a queryable schema. Rather than standing up one MCP server per data source, you give every agent the same runtime; the agent writes standard SQL and Coral translates it into API calls or file reads, handling authentication, pagination, rate limits, and schema mapping along the way. Cross-source JOINs run in a single statement — Coral's own demos join PagerDuty incidents to GitHub pull requests in 84ms and Slack messages to Linear issues in 190ms. It's aimed squarely at engineering teams. Coding-agent users on Claude Code or Codex plug Coral in over MCP (claude mcp add coral -- coral mcp-stdio) or drive it from the CLI, and Coral reports 31% higher retrieval accuracy at 70% lower cost versus wiring data-source MCPs directly in its own benchmark. An AI SRE agent can rank active degradations across Datadog, GitHub, and PagerDuty in one query; a security team can join identity, cloud, and code data into a single access-risk view. The engine leans on your query history too. DESCRIBE EXTENDED exposes columns, relationships, and recommended joins — a PagerDuty service key maps to Datadog service health and GitHub deployments — and Coral uses query patterns, schema access history, and source statistics to sharpen discovery, caching, and execution over time. Query pushdown and caching cut unnecessary API calls and token-heavy tool loops. Because it's a read layer by design, agents inspect production systems without mutating anything upstream, and scoped tokens, workspace isolation, and per-source permissions keep access tight. Against simply pointing an agent at a single provider MCP, Coral earns its keep when you're juggling several sources and someone on the team already writes SQL. For one-off lookups against one system, a direct MCP is less machinery. This is a serious tool for