Debashish Ghosal, a software developer, has developed an open-source governance tool called mcp-fabric-toolmesh to manage Model Context Protocol (MCP) servers. The project originated from the realization that connecting multiple MCP servers to a single AI agent without oversight posed significant security risks, such as granting an agent unauthorized production deployment capabilities alongside routine documentation searches. Ghosal identified that existing solutions like API gateways could not interpret specific MCP tool schemas, leading to the creation of a centralized control plane capable of enforcing OPA (Open Policy Agent) policies, requiring human approval for sensitive actions, and maintaining an audit trail.

The development process involved integrating several planned versions into the initial release, focusing on a robust UI that includes role-based filtering, trust posture color coding, and an approval panel. Building the system required addressing 170 type errors across 41 files, which the author attributed to unclear initial design rather than functional bugs. The final implementation includes 474 tests and manages permissions by normalizing raw tool names into domain-level capabilities, such as mapping diverse promotion commands to a single deployment policy. The system also supports identity tokens per agent class and capability packs for permission bundling.

On July 25, 2026, the project reached version 0.3.0, introducing security hardening and compliance features. This update includes Schema-Digest Mappings, which store a SHA-256 hash of capability mappings to detect and block tool schema drift. The release also added 10 new OPA rules, bringing the total to 31, along with 326 Python tests and 75 Playwright end-to-end testing scenarios. The tool is currently available for installation via pip and is hosted on GitHub to address the growing ecosystem need for standardized AI tool governance.

Debashish Ghosal, a software developer, has developed an open-source governance tool called mcp-fabric-toolmesh to manage Model Context Protocol (MCP) servers. The project originated from the realization that connecting multiple MCP servers to a single AI agent without oversight posed significant security risks, such as granting an agent unauthorized production deployment capabilities alongside routine documentation searches. Ghosal identified that existing solutions like API gateways could not interpret specific MCP tool schemas, leading to the creation of a centralized control plane capable of enforcing OPA (Open Policy Agent) policies, requiring human approval for sensitive actions, and maintaining an audit trail.

The development process involved integrating several planned versions into the initial release, focusing on a robust UI that includes role-based filtering, trust posture color coding, and an approval panel. Building the system required addressing 170 type errors across 41 files, which the author attributed to unclear initial design rather than functional bugs. The final implementation includes 474 tests and manages permissions by normalizing raw tool names into domain-level capabilities, such as mapping diverse promotion commands to a single deployment policy. The system also supports identity tokens per agent class and capability packs for permission bundling.

On July 25, 2026, the project reached version 0.3.0, introducing security hardening and compliance features. This update includes Schema-Digest Mappings, which store a SHA-256 hash of capability mappings to detect and block tool schema drift. The release also added 10 new OPA rules, bringing the total to 31, along with 326 Python tests and 75 Playwright end-to-end testing scenarios. The tool is currently available for installation via pip and is hosted on GitHub to address the growing ecosystem need for standardized AI tool governance.