withcoral/coral

What it solves

Connecting AI agents to multiple external services typically requires managing many individual MCP servers, manually pasting API tokens into each, and overloading the agent's context window with every possible tool definition. Coral consolidates these connections into a single, secure, and policy-governed gateway.

How it works

Coral acts as a centralized MCP server that runs agent-generated TypeScript programs in an isolated, secure sandbox. Instead of loading all tool definitions at once, the agent searches for the specific operations it needs. Coral injects the required credentials, enforces user-defined security policies (such as requiring manual approval for specific actions), and records every execution. Successful programs can be saved as reusable functions, allowing agents to perform complex multi-step tasks more efficiently.

Who it’s for

Developers and teams who use AI agents (like Claude Code, Cursor, or VS Code) to interact with internal or third-party services (GitHub, Jira, Notion, etc.) and need a secure, auditable, and efficient way to manage those integrations.

Highlights

• Reduced Context Usage: Agents only retrieve the specific operations they need, minimizing token consumption and round trips.
• Secure Sandboxing: Programs run in an isolated environment without direct access to secrets or the network, with credentials injected only by Coral.
• Granular Policy Control: Define per-operation rules, including manual approval workflows for sensitive actions.
• Self-Hosted: Full control over your data, credentials, and usage history by running the server on your own infrastructure.
• Extensible: Supports any service with an OpenAPI document or an existing MCP server.