The problem
Not maliciously. It just doesn’t know. It doesn’t know which branch you’re on. It doesn’t know what your Bridge is actually running. It doesn’t know the deploy target you changed last Tuesday. So it invents plausible answers — and you spend twenty minutes finding out why they’re wrong.
That’s not a model problem. It’s an information problem.
One line in your config. From that point, compatible AI clients can ask TARX for project context, tool state, and route-aware evidence instead of guessing.
{
"mcpServers": {
"tarx": {
"url": "https://mcp.tarx.com/mcp"
}
}
}Remote MCP is for discovery and public-safe context. Private files, local memory, protected secrets, and device context stay on the Computer through local MCP.
Use the remote TARX MCP endpoint for compatible clients that need public-safe tool context, install guidance, runtime status, and route-aware evidence. Use local MCP when the client needs private project context, local files, protected secrets, or device state.
Installed TARX can expose local runtime capabilities through local MCP servers: approved context, memory search, task state, code/session state, tool proposals, and evidence. The exact local tool set depends on what is installed, enabled, and approved on that Computer.
MCP can help teams test private-runtime workflows, approved tool access, route truth, and evidence capture before any hosted or high-compute path is enabled.
TARX should show what was read, what was proposed, what needs approval, and what evidence exists. It should not claim autonomous execution or Supercomputer usage without explicit permission and route proof.
TARX starts free on your Computer. Supercomputer is permissioned power when the Computer is not enough. Give compute. Get compute. Buy more only when you need more than you give.