Coral, an open-source gateway for connecting AI agents to business systems, is offering a code-based approach to tool use with policy checks at every provider call. The self-hosted software sits between an agent and services such as GitHub, Slack or Linear, allowing the agent to compose short TypeScript programs and save working programs as reusable tools.

The project is aimed at a recurring problem in agent workflows: models often inspect the same tool descriptions, move large responses through their context and rebuild multi-step procedures for each task. In Coral, an agent can search for available provider operations, write a program that combines them, filter returned data and retain only the result needed for the task. A successful program can then be saved as a private custom tool and called directly later. Workspace owners can publish selected tools for broader team use.

Coral says each program runs in a fresh V8 isolate without network access, imports or embedded credentials. Provider requests are intercepted by the gateway, checked against workspace policy and supplied with credentials only after they pass those checks. Policies can allow an action, block it or require a person's approval. The system also attributes runs to the user or agent that initiated them and records approvals.

That model allows controls to be narrower than simply exposing or hiding an entire provider operation. A custom tool can encode a limited business action—for example, changing one field only on tickets that have been inactive for a defined period—while Coral continues to evaluate each underlying provider request. Read-only operations run immediately by default, according to the repository, while other actions require approval unless administrators change the policy.

The gateway continues server-side execution when an agent disconnects, including workflows paused for human authorization. It can connect through its provider catalog, an OpenAPI document or an MCP server. Agents access those capabilities through Coral's execution tool, first discovering signatures and then submitting the program that performs the work.

The repository describes Coral v1 as a clean break from its earlier SQL interface. Existing 0.x state is not converted, and users are directed to an upgrade guide. The core repository uses the Apache License 2.0, while a separate enterprise directory has its own license. Planned enterprise features include SAML single sign-on, SCIM provisioning, advanced role controls, an immutable audit log and support commitments.

Coral's central proposition is that reusable code can reduce repeated agent work while giving administrators explicit control over sensitive actions. Its effectiveness will depend on deployment and policy design, but the open-source release provides teams with a self-hosted implementation they can inspect and test.