This is in development. Details may change before it is available.
Register an MCP server
Owners and admins open AI Gateway, then MCP servers, and add a remote server by its URL (HTTP or SSE). Choose how Altrum authenticates to it:
Servers that run as local processes are not supported. Altrum refuses addresses on private networks, and checks the address again every time it connects, so a name that later points somewhere private is refused too. Code mode, running tools without your application asking, and skipping the server’s certificate check are not offered.
Connect your own account
For servers that act as each person, open AI Gateway, then MCP servers, and connect your account. Tool calls made with your personal key then run as you. A request made with an organization or AI system key cannot use a per-person server.Choose who can use which tools
Owners and admins choose which servers and tools are allowed for the organization’s keys, for each AI system, and for each key. A key sees only the tools it is allowed. A list set on an AI system limits every key of that AI system, and a list set on the organization limits the organization’s own keys. A server that is not named in a list is not reachable, and* stands for every tool of a server.
Use the tools
- Coding agents
- Your applications
Add Altrum’s MCP endpoint,
https://api.altrum.ai/mcp, to your coding agent with your personal key. The agent sees the tools your key allows, across all your organization’s servers. See Coding agents.When a tool cannot run
A server that works with each person’s own account needs a personal key. Two answers explain what is missing, word for word:<server> needs a personal key: it works with each person's own account, which a shared key does not have. Use a personal key to call it.The tool is also left out of the list an organization or AI system key sees.Your account is not connected to <server>. Connect it in Altrum under AI Gateway, MCP, then try again.
Governance on every call
- Tool arguments are checked by your guardrails before the tool runs, and results before they return to the model. See Guardrails at runtime.
- Every call is logged: the tool, the server, who called it, how long it took and whether it worked. Arguments and results are kept only when content logging is on. See Request logs.