Agent-Fabric documentation

Agent-Fabric publishes the models, tools and agents already running on your own hardware as private services your team — and your customers — can call. This is the reference for the fabric CLI, the control-plane API and the SDKs.

New here?Install and publish your first service in about five minutes —Quickstart

• Get running

  • QuickstartInstall, sign in, connect, publish
  • Connect a machineJoin a laptop, server or GPU box to your network
  • Publish a serviceExpose a local model, MCP tool or internal API

• Share access

  • Grant scoped accessMint a capability token for one action, with an expiry
  • Give a customer a demoTime-boxed access you can revoke in one click
  • Run agents locallyModel runtimes and session-backed agent loops

• Understand and fix

  • How it worksWhat runs where, and what the cloud never sees
  • TroubleshootingWhen a machine is offline or a name will not resolve
  • See it runReal recordings of the CLI and both consoles

• Reference

  • CLI referenceEvery fabric command and flag
  • HTTP APIControl-plane routes, auth tiers, entitlements
  • SDKsGo, Python and TypeScript clients

What the control plane can and cannot see

Your agents, models and services run on your own machines. Agent-Fabric gives them a private encrypted mesh (WireGuard) and a control plane that brokers keys, device identity and policy.

The control plane sees names, health and usage counts. It does not see prompts, model outputs or any traffic between your machines — those are end-to-end encrypted, with a relay used only when two machines cannot reach each other directly, and the relay carries ciphertext it cannot read.

If you only read one more page, read How it works.