Pair local compute with your account

Updated 13 October 2026
Auto entry flow checked against Console v0.2.8. Available actions depend on project access and runtime configuration.

Start with Auto on a supported Mac#

In Console, open Settings → Organisation → Compute → Pair local compute. The default dialog offers Auto; Advanced CLI setup is a separate option.

1. Choose Download Auto and install the companion. The dialog currently lists Apple silicon, macOS 26.2 or later, and at least 16 GB of memory.
2. Choose Open Auto to start the device connection.
3. Approve your project in Console, then keep Auto open to stay connected.
4. Return to Compute to check online status, then assign a compatible model and verify readiness.

Opening Auto or approving pairing does not prove a model is ready. If Auto reports a setup failure, read that error and check device eligibility and connectivity before retrying. Do not assume the device connected successfully.

Advanced CLI setup: what you need#

• A compatible HybrIE runtime already installed and running on the machine. Use the runtime’s setup instructions or ask your operator for the supported build and model requirements.
• A current Stimulir CLI with compute pair, and permission to approve a device in your Console organisation and project.
• A runtime base address reachable from the Console backend. For production Console, 127.0.0.1 points at the backend itself, not your laptop.
• Enough real hardware capacity and the required model files. A pairing record does not install a runtime or make unsupported hardware compatible.

Use an operator-configured private network connection that the backend can actually reach. Both the pairing dialog and released CLI require a private IP address or localhost; public hostnames, public IPs, and ordinary public tunnel URLs are rejected. A private address that only opens on your laptop or home network is insufficient. If no routed private connection exists, ask your operator to establish one before pairing. Restrict access to the runtime and preserve its required authentication; do not expose an unauthenticated service to make pairing work.

1. Start CLI pairing on the machine#

In Pair local compute, expand Advanced CLI setup. Enter the backend-reachable HybrIE runtime address, without /v1. The dialog generates a command to run on that machine.

stimulir compute pair \
  --api-base https://api.stimulir.com \
  --runtime-url http://10.0.0.20:8011 \
  --backend auto

For an Apple Silicon Metal runtime, use --backend metal. This declares the intended backend; the runtime must still report genuine compatible capabilities.

If local capability discovery uses a different address, fill Runtime address on this machine (optional), or add --probe-url http://127.0.0.1:YOUR_PORT. The probe address is for the CLI; the registered runtime address must remain reachable by the backend.

2. Review the code in Console#

1. Read the pairing code and verification link printed by the CLI.
2. Enter that code in Pairing code from the machine and select Review pairing, or open the printed verification link.
3. Check that the code and device belong to your machine. Select the intended organisation/project and approve access.
4. Return to Compute and check the device record and online status.

Pairing grants device access. Keep the code and resulting credentials private. The CLI sends a heartbeat every 30 seconds; keep both it and HybrIE running. --once stops after registration and does not keep the machine online.

3. Assign a compatible model#

1. Open Models and select the model you want to serve.
2. Choose the paired compute option compatible with that model.
3. Select Assign compute and enable. Inspect setup status until it reports Ready, or read the reported failure.
4. Run an authorised small model test, then inspect the trace and runtime evidence to verify where the request executed.

Paired, Online, and Ready describe different things. A device can be registered but offline, or online without a worker able to serve the model.

Clef and Clef Flash on Metal

The current workers require CUDA SGLang or a verified native MLX Metal runtime. A Metal worker must advertise decision_runtime_backend=metal, decision_runtime_joint_heads=true, and have Cloudflare/clef or Cloudflare/clef-flash installed. A declared memory size does not establish those capabilities. NPU pairing alone does not support the current Clef workers.

Private pairing and shared peer compute#

Your paired machine starts private to your workspace. In Compute → Shared peer compute, Allow admin sharing is a separate owner decision. It permits an administrator to put the machine in a shared pool; consent alone does not grant other businesses access.

An administrator must configure pool membership and the applicable business access grants. Online status, runtime reachability, and approved inference admission still apply. Review the sharing notice and recipients before consenting. Revoking sharing blocks use through those pools.

Peer sharing is not monetised. Pairing does not rent a cloud GPU, but it uses your machine’s resources and may require model downloads. Separate cloud inference, judges, and training can still incur charges.

If the machine cannot serve a model#

Auto is the default companion flow for supported Macs. The advanced CLI flow remains available for an existing compatible runtime. See Pairing reference, Workers, and Tunnels for the related contracts.