Integrate anything. Control it with your own AI.
MCP Universe is the connector layer under Synapse. It turns a system you already run — a CRM, a billing stack, an ERP, a database, a switch — into a governed MCP surface: every operation classified by risk, every call authorised and recorded, every credential injected at the edge so that no model ever sees it. Your CEO agent and your teams use the result as ordinary tools.
Every connector says how it was built.
A connector is a declared manifest, not a prompt. It states which tier it came from and how much of it was inferred, so you know what to trust before you grant it anything.
An existing MCP server
Already have an MCP server? Put it behind the gateway and it inherits the policy engine, approvals and audit log. About a minute.
OpenAPI, GraphQL, gRPC
Upload the spec on the Request a connector page. The manifest is compiled from the machine-readable description; ten to thirty minutes, most of it review.
Human documentation only
No spec, just docs. An AI engine you configure — a local vLLM model or your own provider key — drafts the manifest. Every inferred operation is marked as such.
A protocol, not an API
SSH, SNMP, Modbus, SQL. Driven through a restricted command template whose rendered output is re-validated in full before it runs.
A user interface and nothing else
Experimental, always. We list it so you know where the ladder ends.
Pipedrive · Jira Data Center
Connectors published for every company on the server to install. Publishing is a second decision after a request is approved; a bad version can be quarantined with a kill switch and installations already using it keep working.
Your AI gets a tool. It does not get your password.
- Credentials never reach a model. They are stored once, in the console, and injected at the gateway when a call is made. A credential in a tool result is a compile error in the codebase, not a runtime check.
- Every operation carries a risk class. Read-only installs are the default. Writes above the grant you gave escalate to a pending approval that is decided in your Synapse cabinet, in the same place as CEO approvals.
- Per-company endpoints. Each company gets its own
/t/<company>/mcpsurface; tools are projected per principal, so a caller only sees what it is allowed to use. - An audit log of every call — who, which connector, which operation, allowed, denied or escalated.
- Everything is done from the web. Credentials, AI engines, installations, approvals and the audit log are console pages. No terminal steps to run an integration.
What the CEO sees, and what we have verified.
Settings → Connectors
Browse the catalogue, connect a connector to your company and decide what it may do. An approval card appears in CEO chat when a connector operation needs you. Built and merged in September; not yet exercised with a customer company in production.
Connector tools in the loop
The CEO carries a constant set of seven mcpu__* meta-tools and discovers connected systems through them, so adding a connector never changes its prompt and never breaks its cache. Same qualification: merged and covered by the test suite, not yet verified with a live tenant.
mcp.catalation.com
The server, the console, the catalogue with Pipedrive and Jira, credential and engine management and the operator API have been in production since 30 September 2026. The public endpoint is https://mcp.catalation.com/mcp.
MCP Universe is a separate, self-hostable product of Catalation. Synapse is its first consumer, not its owner — your keys, your data, on your own server if you prefer.