Turn-key text inference worker for AI Power Grid. Run a local model, connect to the Grid, and serve paid inference jobs.
Start at the verified operator onboarding page. It checks the immutable release envelope before exposing a platform download, shows current model demand and worker redundancy, and provides a public online worker check. Linux operators who want to add independent capacity can join the open text-worker cohort.
The underlying immutable assets are also available from Releases:
Release candidates include SHA256SUMS, worker-release.json, an SPDX SBOM,
the checksum-verifying Linux installer, and GitHub build-provenance
attestations. The candidate manifest records
macOS and Windows signing state explicitly. Every release requires supervised
staging; platform signing is recommended but optional. Unsigned builds are
identified in the manifest and release notes so operators can make an informed
choice. Published releases are immutable; corrections are issued as a new
version rather than replacing a tag or binary.
The current public release is v0.3.8. Ordinary single-model workers use the
native Console enrollment flow. The candidate Grid key stays on the worker and
becomes active only after a signed, short-lived Console approval. Advanced
multi-backend or parallel operators continue to use a scoped Grid API key.
Source on main may contain a newer draft candidate. Use the version exposed
by the verified /run release gate unless you are performing supervised
candidate qualification.
| Platform | File |
|---|---|
| Windows x64 | grid-inference-worker-windows-x64.exe |
| macOS ARM64 | grid-inference-worker-macos-arm64.zip |
| Linux x64 (Ubuntu 22.04+) | grid-inference-worker-linux-x64 |
| Linux ARM64 (Ubuntu 24.04+) | grid-inference-worker-linux-arm64 |
Windows — Double-click the exe. A setup wizard opens in your browser at http://localhost:7861.
macOS — Unzip, then open Grid Inference Worker.app.
Linux — chmod +x grid-inference-worker-linux-x64 && ./grid-inference-worker-linux-x64
Beginning with v0.3.7, Linux operators can download
install-worker.sh from the same release, inspect it, and run it. The script
selects x64 or ARM64, verifies the binary and release manifest against
SHA256SUMS, and installs to ~/.local/bin without starting the worker or
asking for credentials. It is not a curl | sh installer.
No Python or dependencies needed. Just install a backend (Ollama is easiest), run the worker, and follow the wizard. Connect Grid account is the normal single-model path: the worker opens Console for human approval and installs a worker-only credential without returning the generated key to the browser. Advanced multi-backend or parallel operators use an existing scoped key from the developer Console because each enrolled credential is bound to one exact worker name.
The normal Console-enrolled flow confirms registration and then asks Core to route a randomized, hard-targeted connectivity canary through that exact worker before reporting success. The canary has no charge, den, payout, strike, or validator effect; it proves the route and exact response, not model identity or general quality. Advanced account-key setups receive registration-only confirmation because those keys are not bound to one exact worker. A running process alone is not an online worker. The dashboard distinguishes connecting, partially connected, online, and unavailable status. If setup cannot confirm a connection, open Logs and check the backend, API key, and worker name.
Payouts use the account that owns the API key. Native enrollment uses the
account that approved the worker credential. Manage the payout wallet in
console settings, not in the
worker. The legacy local WALLET_ADDRESS setting does not set a Grid payout
destination. Do not enter a wallet private key in this application.
Den is the Grid's work-accounting unit for accepted jobs, not a token amount or
fixed exchange rate. Demand, routing, measured work, availability, competition,
and settlement all affect results. Use the live workload and payout evidence on
/run; it is historical evidence, not an earnings
or ROI guarantee.
The desktop manager can copy an authenticated dashboard link for another local
browser. In headless mode, run grid-inference-worker --show-dashboard-link
explicitly; normal startup logs never print the dashboard token.
Once your worker is running, chat with your model at aipg.chat — select your model in the upper selector.
The v0.3.9 candidate adds a model roster to setup and the dashboard. Console enrollment remains limited to one model/connection; multiple models require an advanced account key. Settings edits each roster entry's endpoint, engine, credential, and concurrency independently. Stored credentials are never shown and are not copied to a changed endpoint.
GRID_BACKENDS entries may set paused (boolean), max_context (zero for
detection, otherwise a ceiling on detected capacity), and schedule (JSON
windows). An omitted schedule inherits GRID_SCHEDULE; an explicit empty one
uses the entry's normal concurrency all week. No selected days means an explicit
all-week pause. Custom multi-window schedules remain editable as JSON, and
explicit modalities/vision declarations survive roster edits. Context limits
advertised to Grid do not configure backend VRAM allocation.
Already run persistent infrastructure? AI Power Grid is recruiting two more unrelated Linux/systemd or persistent Docker operators for the initial validator cohort. The validator is CPU-only: it does not require a GPU, stake, or a worker, and preview validators have no routing, reward, strike, or slashing authority.
Each candidate must complete a 72-hour evidence window before it can qualify.
Multiple nodes controlled by the same person or organization count as one
independent operator, including when that operator also runs Grid workers.
Start at aipowergrid.io/validate, then post
only your public val_* status ID in the
qualification cohort issue.
Never post an API key or private key.
Override config from the command line. The web dashboard is always available at http://localhost:7861 regardless of how you start the worker.
grid-inference-worker \
--model llama3.2:3b \
--backend-url http://127.0.0.1:11434 \
--api-key YOUR_API_KEY \
--worker-name my-worker--model NAME Model name (e.g. llama3.2:3b)
--backend-url URL Backend URL (e.g. http://127.0.0.1:11434)
--api-key KEY Grid API key
--worker-name NAME Worker name on the grid
--port PORT Web dashboard port (default: 7861)
--host HOST Dashboard bind host (default: 127.0.0.1)
--gui Show the desktop control window (default for binaries)
--no-gui Skip the desktop control window
--install-service Install as a system service (auto-start on boot)
--uninstall-service Remove the system service
--service-status Check if the service is installed
Copy .env.example to .env and fill in your values, or configure through the web setup wizard.
| Variable | Default | Description |
|---|---|---|
GRID_API_KEY |
Scoped Grid API key (create one); advanced path for multi-backend or parallel operators | |
GRID_ENROLLED_WORKER_NAME |
Exact worker name installed by secure Console enrollment; do not set manually | |
MODEL_NAME |
Model to serve (e.g. llama3.2:3b) |
|
BACKEND_TYPE |
ollama |
ollama or openai |
OLLAMA_URL |
http://127.0.0.1:11434 |
Ollama endpoint |
OPENAI_URL |
http://127.0.0.1:8000/v1 |
OpenAI-compatible endpoint (vLLM, SGLang, etc.) |
OPENAI_API_KEY |
API key for OpenAI-compatible backend | |
GRID_WORKER_NAME |
Text-Inference-Worker |
Worker name on the grid |
GRID_MAX_LENGTH |
32768 |
Fallback output-token budget when the request omits one |
GRID_MAX_CONTEXT_LENGTH |
131072 |
Maximum advertised context window (auto-detected when possible) |
GRID_NSFW |
true |
Accept NSFW jobs |
WALLET_ADDRESS |
Legacy local value; does not control Grid payouts |
Requires Python 3.11+.
pip install -e .
grid-inference-workerOn Windows you can also use:
.\scripts\run.ps1cp .env.example .env
# Edit .env with your values
docker compose up -dThe dashboard is available at http://localhost:7861 and binds to loopback by
default. Use --host 0.0.0.0 only when you deliberately need LAN access; the
generated dashboard token is still required.
Run the worker on boot without needing to stay logged in. Works on Windows (startup registry), Linux (systemd), and macOS (launchd).
# Configure the worker first (run it once to set up .env), then:
grid-inference-worker --install-service
# Check status
grid-inference-worker --service-status
# Remove
grid-inference-worker --uninstall-service| Backend | Type | Setup |
|---|---|---|
| Ollama | ollama |
Install Ollama, ollama pull llama3.2:3b, done |
| LM Studio | ollama |
Load a model, enable server in Developer tab |
| vLLM | openai |
--served-model-name + set OPENAI_URL |
| SGLang | openai |
Point OPENAI_URL at SGLang's OpenAI endpoint |
| LMDeploy | openai |
lmdeploy serve api_server + set OPENAI_URL |
| KoboldCpp | openai |
Enable OpenAI-compatible endpoint |
Ollama is the easiest way to get started. The setup wizard auto-detects it and lets you pick a model.
For any backend that exposes an OpenAI-compatible API (/v1/chat/completions), set BACKEND_TYPE=openai and point OPENAI_URL at it.
For high-performance inference with vLLM, see our detailed guides:
- vLLM Setup Guide - Installation, configuration, and integration
- vLLM Optimization Guide - Performance tuning, benchmarking, and production best practices