Metorial

As companies let more teams connect AI agents to real company apps, governance breaks down fast. People install MCP servers on their own laptops, so nobody knows which tools reach company data. Personal API tokens get pasted into local config files and keep working for months with no owner or expiry. When an agent changes the wrong record, there is often no way to tell whose access it used, because a shared credential leaves no record of what actually ran.

Metorial puts one governed layer between the AI clients a company already uses (Claude, ChatGPT, Cursor, Codex) and the tools those agents reach, connecting to over 1,100 MCP servers and 20,000+ tools from one place. People sign in once through their existing SSO (SAML, OIDC, SCIM), and their groups and roles determine which integrations and skills they can see, down to single tools or read-only sets. Each person gets a single 'Magic MCP URL' to paste into any supported client, so approved tools follow them everywhere without re-configuring access. Every tool call is screened by 'Protoguard' for prompt injection and schema changes, runs isolated in its own enclave using that person's own credentials (via Metorial Vault), and is logged to an audit trail. Teams can also write shared 'Skills', reusable agent workflows with instructions and tool references, so an expert's process becomes available to every teammate's agent. The product is open core, available on GitHub, and can be self-hosted, run on-prem/air-gapped, or used via Metorial Cloud.