Skip to main content
OpenGrid
Provider guide

Add your compute to OpenGrid

Connect a standalone machine, or review the planned Dynamo-managed cluster path. OpenGrid coordinates across provider boundaries; each provider owns its local trust and execution boundary.

Preview guided setup View node downloads
Closed developer preview. The onboarding flow is implemented, but public provider enrollment and signed, notarized macOS downloads are not open yet. These steps describe the intended provider experience once the release channel is available.

Choose the provider boundary

Standalone path

Mac or single host

The current curated node path supervises llama.cpp and applies local resource limits. MLX support exists as an adapter, but curated MLX distribution is not available.

Compatibility client only

Dynamo cluster

A planned provider connector will represent the deployment. Dynamo will own its internal GPUs, KV routing, prefill/decode pools, and autoscaling.

Current standalone placement uses trust, region, observed performance, load, and availability. SLA- and price-aware marketplace placement is planned. OpenGrid will never schedule an individual Dynamo worker or transfer KV state between providers.

Before connecting a standalone node

Supported hostApple Silicon with macOS 14+, or a dedicated 64-bit Linux or Windows host, with at least 16 GB memory.
Working spaceKeep at least 25 GB free for a verified model, download staging, and runtime state.
Stable connectionAllow outbound network access. A provider node does not need an inbound public port.

Install the app or CLI before creating a token: an enrollment token is single-use and expires after 15 minutes.

Connect a provider machine in seven steps

  1. Create and activate your provider account

    Sign in and activate the provider profile. During activation, acknowledge that a serving machine processes prompts and generated output in plaintext memory.

  2. Install the node software

    Install the signed macOS app when its release channel opens, or install the OpenGrid CLI and packaged llama-server runtime. Complete this before generating a token.

    Open Provider downloads

  3. Run the preflight checks

    Confirm the configuration, runtime, coordinator connection, and local daemon state.

    opengrid node doctor

  4. Generate a one-time token

    Open Provider → Nodes → Add a node, choose macOS, and select a resource profile. Generate the token only when you are ready to join; it expires after 15 minutes and enrolls one device.

  5. Join with safe local defaults

    Start a MacBook with Night, or a dedicated host with Server. Read the token without echoing it so it stays out of shell history and the join arguments.

    printf 'Enrollment token: '
read -r -s OPENGRID_ENROLLMENT_TOKEN
printf '
'
export OPENGRID_ENROLLMENT_TOKEN
opengrid node join --profile night

    join remains in the foreground. It creates the Ed25519 device identity locally, enrolls over an outbound secure tunnel, and starts the verified starter-model download.

  6. Verify readiness in another terminal

    Watch enrollment, connection, model download, and runtime state while the join process continues.

    opengrid node status --json
opengrid node models downloads
opengrid node models list
opengrid node logs --follow

  7. Keep it running safely

    Enable the native app at login or install the shipped launchd service. Preserve ~/.opengrid; it contains the node identity, certificates, model state, and unsettled receipts.

A node is ready only when all four checks pass

  • The device is enrolled.
  • The coordinator connection is active.
  • The local inference runtime is available.
  • At least one verified model is installed and loaded.

Choose how your machine contributes

Recommended for MacBooks

Night

Runs from 23:00–07:00 and pauses on battery or user activity, with conservative thermal limits.

Everyday computer

Balanced

Available throughout the day with safe headroom, battery protection, and activity-aware pauses.

Dedicated hardware

Server

Higher utilization for an always-on machine with stable power, cooling, storage, and network access.

The daemon enforces every profile locally. The coordinator can offer work, but it cannot raise your limits. Advanced operators can tune individual controls with opengrid node limits set.

Security and local state

Plaintext boundary: a provider machine sees the prompt and generated output in plaintext memory while performing inference. Do not run a node on a machine or network where you are not authorized to process that data.
Typed inference only: the node never executes consumer tools, shell commands, model-supplied code, or arbitrary binaries.

The private identity key never leaves the device. Protect ~/.opengrid, run the daemon as an unprivileged user, keep security updates enabled, and expose no node ports publicly.

Availability does not guarantee assignments or earnings. OpenGrid points are off-chain accounting units, not money or a promise of redemption.

Dynamo cluster providers

Compatibility foundation only. Dynamo cluster enrollment is not enabled for billable traffic yet. Do not expose a Dynamo frontend publicly or treat its usage response as settlement evidence.

The first integration profile targets a pinned NVIDIA Dynamo release with vLLM on Kubernetes. The provider connector will run beside the cluster, connect outbound to OpenGrid, call the Dynamo frontend over a private network, and sign deployment-level usage evidence.

  1. Operate a supported Linux/NVIDIA Dynamo deployment in your own trust domain.
  2. Install the forthcoming OpenGrid connector with an immutable deployment fingerprint.
  3. Enroll the connector, then advertise models, region, SLA headroom, and offer terms.
  4. Pass health, feature, provenance, cancellation, and synthetic canary checks.
  5. Run shadow requests before OpenGrid enables settlement for the deployment.

MacBooks remain on the standalone path: Dynamo's production support matrix targets Linux and NVIDIA GPUs, not macOS or Apple Silicon.

Advanced: dedicated Docker host

Container release unavailable. The production image channel is not currently published. Do not copy an unpinned :latest example; wait until Provider downloads supplies a reviewed immutable image digest.

Docker is intended for a dedicated Linux or NVIDIA host, not the normal MacBook path. When a signed release is available, pass the one-time token from a secret store, drop unnecessary capabilities, keep the state volume persistent, and use the Server profile.

docker run -d --name opengrid-node \
  --gpus all \
  --restart unless-stopped \
  -v opengrid-data:/var/lib/opengrid \
  -e OPENGRID_ENROLLMENT_TOKEN \
  "$OPENGRID_NODE_IMAGE" node join --profile server

After release, pin OPENGRID_NODE_IMAGE to the published digest and remove the enrollment token after the first successful start. The in-container health endpoint is http://127.0.0.1:7332/healthz.

Ready when the provider channel opens

Install first, generate the short-lived token last, and begin with Night on a MacBook or Server on dedicated hardware.

Preview Provider nodes