obolnetwork/monetize-guide

Monetize Guide

Step-by-step guide to expose local GPU resources or HTTP services as x402 payment-gated endpoints with on-chain discovery via ERC-8004.

When to Use

• User says "monetize my machine", "sell my GPU", or "expose my service for payments"
• Setting up a new paid endpoint from scratch
• Need to figure out what to sell and at what price

When NOT to Use

• Managing existing offers (list/status/delete) — use sell directly
• Buying inference from others — use buy-x402
• Cluster diagnostics — use obol-stack

Workflow

Follow these phases in order. Stop and ask the user for confirmation before executing phase 4.

Phase 1: Pre-flight Checks

Verify the environment is ready before proceeding.

# 1. Check cluster is running
obol kubectl get nodes

# 2. Check the agent is initialized (has RBAC for monetization)
obol kubectl get clusterrolebinding openclaw-monetize-read-binding -o jsonpath='{.subjects}'

# 3. Get the wallet address (auto-generated by agent)
python3 ${OBOL_SKILLS_DIR:-/data/.openclaw/skills}/ethereum-local-wallet/scripts/signer.py accounts

If cluster is not running: Tell the user to run obol stack up first. Do NOT run it yourself — it takes several minutes and changes system state.

If agent has no RBAC subjects: Tell the user to run obol agent init.

If no wallet address: The wallet is created during obol stack up. If missing, the stack may not have completed setup.

Phase 2: Detect What Can Be Sold

For GPU / Inference

# Check what Ollama models are available locally
curl -s http://localhost:11434/api/tags | python3 -c "
import json, sys
data = json.load(sys.stdin)
for m in data.get('models', []):
    size_gb = m.get('size', 0) / 1e9
    print(f\"  {m['name']:30s} {size_gb:.1f} GB\")
"

Report the available models to the user. If no models are found, suggest they pull one:

ollama pull qwen3.5:4b    # Smallest current Qwen, ~3.4 GB (low-RAM laptops)
ollama pull qwen3.5:9b    # Validated baseline, ~6.6 GB
ollama pull qwen3.6:27b   # High quality, ~17 GB (needs ≥32GB RAM)

For External LAN Resources (GPU servers, vLLM, etc.)

Ask the user if they have a GPU server or inference endpoint on their local network. If yes, get:

• The endpoint URL (e.g., http://192.168.0.202:8000/v1)
• The model name served at that endpoint

Verify it's reachable and OpenAI-compatible:

# Probe the external endpoint
curl -s <ENDPOINT_URL>/models | python3 -c "
import json, sys
data = json.load(sys.stdin)
for m in data.get('data', []):
    print(f\"  {m['id']}\")
"

If reachable, this endpoint can be sold by bridging it through LiteLLM (see Phase 4).

For HTTP Services (in-cluster)

# List services in the cluster that could be monetized
obol kubectl get svc -A --no-headers | grep -v 'kube-system\|traefik\|x402\|monitoring\|erpc\|obol-frontend'

Ask the user which service they want to expose and on which port.

Phase 3: Research Pricing

Query the ERC-8004 registry to see what comparable services charge.

# Search for registered agents on Base Sepolia
python3 ${OBOL_SKILLS_DIR:-/data/.openclaw/skills}/discovery/scripts/discovery.py search --limit 10

# For each agent with x402Support, fetch their registration to see pricing
python3 ${OBOL_SKILLS_DIR:-/data/.openclaw/skills}/discovery/scripts/discovery.py uri <agent_id>

Pricing guidelines (present these to the user with your research):

| Service Type | Typical Range | Notes | |-------------|---------------|-------| | LLM inference (small, <4B) | 0.0005–0.002 USDC/req | Fast, low compute | | LLM inference (medium, 4-14B) | 0.001–0.005 USDC/req | Good quality/cost balance | | LLM inference (large, >14B) | 0.005–0.02 USDC/req | High quality, slower | | Data API / indexer | 0.0001–0.001 USDC/req | Depends on query complexity | | Compute-heavy (GPU hours) | 0.10–1.00 USDC/hour | Fine-tuning, training |

Always present your research and recommendation to the user and ask them to confirm the price before proceeding.

Phase 4: Sell the Service

Only proceed after the user has confirmed the price.

Inference (Ollama model)

obol sell inference <name> \
  --model <model_name> \
  --price <confirmed_price> \
  --register-name "<descriptive name>" \
  --register-description "<what the model does>" \
  --register-skills natural_language_processing/natural_language_generation/text_completion \
  --register-domains technology/data_science

The --wallet and --chain will be auto-resolved (remote-signer wallet, base-sepolia default).

External LAN Resource (GPU server, vLLM, etc.)

Two steps: first bridge the endpoint into LiteLLM, then sell LiteLLM.

# Step A: Add the external endpoint to LiteLLM
obol model setup custom \
  --endpoint <full_url_with_v1> \
  --model "<model_name_at_endpoint>"

# Step B: Sell it through LiteLLM
obol sell http <name> \
  --upstream litellm \
  --port 4000 \
  --namespace llm \
  --per-request <confirmed_price> \
  --wallet <wallet_address> \
  --chain base-sepolia \
  --health-path /health/liveliness \
  --register-name "<descriptive name>" \
  --register-description "<what the service does>" \
  --register-skills natural_language_processing/natural_language_generation/text_completion \
  --register-domains technology/data_science

The --endpoint must include /v1 if the upstream is an OpenAI-compatible server (vLLM, TGI, etc.) — LiteLLM does not append it automatically.

LAN IPs (e.g., http://192.168.0.202:8000/v1) are reachable from inside the k3d cluster without any additional network configuration.

HTTP Service (in-cluster)

obol sell http <name> \
  --upstream <service_name> \
  --namespace <namespace> \
  --port <port> \
  --per-request <confirmed_price> \
  --chain base-sepolia \
  --health-path <health_endpoint> \
  --register-name "<descriptive name>" \
  --register-description "<what the service does>" \
  --register-skills <oasf_skill_path> \
  --register-domains <oasf_domain_path>

Phase 5: Wait for Reconciliation

obol sell http now registers by default. Use --no-register only for local or private-only flows where on-chain discovery is intentionally skipped.

The agent reconciler automatically processes the ServiceOffer through 6 stages.

# Watch the conditions progress (check every 15 seconds, up to 2 minutes)
for i in $(seq 1 8); do
  echo "--- Attempt $i ---"
  obol sell status <name> -n <namespace> 2>&1 | grep -A1 'type:\|status:\|reason:\|message:'
  sleep 15
done

Expected progression:

1. ModelReady → True (instant for HTTP, may take minutes if pulling a model)
2. UpstreamHealthy → True
3. PaymentGateReady → True
4. RoutePublished → True
5. Registered → True (best-effort, non-blocking)
6. Ready → True

If a stage is stuck, check:

# Agent logs for reconciliation errors
obol kubectl logs -l app.kubernetes.io/name=hermes -n hermes-obol-agent --tail=50

# x402-verifier is running
obol kubectl get pods -n x402

Phase 6: Verify the Endpoint

# Get the tunnel URL
obol tunnel status

# Test the endpoint is live and payment-gated (should return 402 with pricing)
obol sell test <name> -n <namespace>

# Or manually:
TUNNEL_URL=$(obol tunnel status 2>&1 | grep -o 'https://[^ ]*')
curl -s -o /dev/null -w "%{http_code}" "$TUNNEL_URL/services/<name>/health"
# Expected: 402

# Inspect the 402 pricing response
curl -s "$TUNNEL_URL/services/<name>/health" | python3 -m json.tool

A 402 response with x402Version: 1 and an accepts array confirms the endpoint is live and payment-gated.

Phase 7: Report to User

Present a summary:

Service monetized successfully!

  Name:       <name>
  Model:      <model> (if inference)
  Price:      <price> USDC per request
  Chain:      base-sepolia
  Wallet:     <wallet_address>
  Endpoint:   <tunnel_url>/services/<name>/v1/chat/completions
  Registry:   Agent #<id> on ERC-8004 (Base Sepolia)

Buyers can discover this service at:
  <tunnel_url>/.well-known/agent-registration.json

To check status:  obol sell status <name> -n <namespace>
To stop selling:  obol sell stop <name> -n <namespace>
To delete:        obol sell delete <name> -n <namespace>

OASF Skills & Domains Reference

Use these when registering for on-chain discovery:

Common skills:

• natural_language_processing/natural_language_generation/text_completion — LLM chat
• natural_language_processing/text_generation — general text generation
• data_management/indexing — data indexing services
• data_management/search — search services
• devops_mlops/model_versioning — training/fine-tuning

Common domains:

• technology/data_science — AI services
• research_and_development/scientific_research — ML research
• technology/blockchain — blockchain data services

Constraints

• Always confirm pricing with the user — never set a price autonomously
• Do NOT run obol stack up without explicit user request — it's a long-running infra change
• Do NOT run obol stack down or obol stack purge — destructive operations
• Registration is best-effort — services become Ready even if on-chain mint fails (wallet may lack gas funds)
• Tunnel URL changes on restart — registration docs auto-update on next reconcile