Configuration, not a project.
Budgets that stop overruns before the spend.
Seats and credit allocations, per-model rate cards, quotas enforced before the model call, append-only ledgers — and a budget guard that pauses requests at 100%.
A switch, a ceiling and a list — for the assistants your people already use.
Claude, ChatGPT and Claude Code can connect to your organisation over the MCP endpoint. Whether they may, and how far, is decided here — off for every organisation until an administrator says otherwise.
The capabilities, stated plainly.
Every row ships today and traces to a capability described on this page.
| Capability | xMatix AI Studio |
|---|---|
| Agents built in the setup screens | Native |
| Template gallery, kept current | Native |
| Bring-your-own model keys with failover | Native |
| Custom tools bundled as capabilities | Native |
| Knowledge bases with automatic ingestion | Native |
| Voice agents on inbound lines | Native |
| Budgeted, step-capped, audited runs | Native |
| Per-model rate cards and quotas | Native |
| Budget guard that pauses at 100% | Native |
| MCP server for external AI clients | Native |
The questions evaluators ask.
Which models can we use?
The ones you choose. Connect your own model endpoints with encrypted keys, mix providers, and order them in a failover chain that switches before the first token. Models pin per agent, or complex questions escalate automatically to a designated model. The platform does not lock you to one vendor's model — your keys, your models, your failover.
Can we cap what AI costs?
Yes — before the spend, not after it. Seats and credit allocations set who may consume, per-model rate cards price the consumption, and quotas are enforced before the model call is made. Usage lands in append-only ledgers, and a budget guard pauses requests the moment a budget hits 100%. The overrun conversation happens in configuration, not in the invoice.
Can we restrict which tools an agent may call?
Yes — tools bundle into named capabilities that you grant like permission sets. Script tools written in the platform, connected tools from integration servers and platform tools are all scoped the same way, so an agent holds exactly the verbs you gave it and nothing more. Tenants mint their own capabilities; the platform's catalog is the floor, not the ceiling.
Do agents run with their own identity or the user's?
The user's. Agents are audience-scoped and run as the requesting user, with that person's permissions and a full audit trail — the collections agent answering a finance analyst sees what the analyst may see. Long-running agent workflows queue as jobs with checkpoints, and every run is budgeted, step-capped and audited, so autonomy stays within limits you set.