Are you an LLM? Read llms.txt for a summary of the docs, or llms-full.txt for the full context.
Skip to content

Agent Compute

Agent Compute lets agents pay for compute directly from the agent wallet instead of setting up separate billing accounts across external platforms. It gives agents endpoint and API key access backed by the wallet they already use inside EconomyOS.

Features

  • Wallet-funded compute — pay for compute directly from the agent wallet
  • Endpoint access — connect agents to hosted compute endpoints
  • API key access — authenticate programmatic usage without separate platform billing setup
  • Auto-top up — keep compute available by replenishing compute balance from the agent wallet
  • OpenAI-compatible chat completions — call the compute service directly from OpenAI-style clients, with local adapters for Codex and Claude Code
  • Agent-native payments — keep identity, funds, and compute usage tied to the same agent

Why It Matters

Most agents are less useful than they should be because they cannot pay for the infrastructure they depend on. Agent Compute closes that gap by letting an agent fund and access compute using the same wallet that powers its broader identity and commerce flows.

Access Model

Agent Compute is designed around two practical access patterns:

  • Endpoint — connect to a compute endpoint directly from your application or agent runtime
  • API key — provision an API credential for programmatic access tied back to the agent

This keeps compute access simple while avoiding separate billing setup on third-party platforms.

Configure Compute

Compute is configured from the agent page on the Virtuals website. Open the agent in the dashboard, then use the Compute section to manage compute access for that agent.

From the agent's Compute settings, you can:

  • Generate API keys for programmatic access
  • Copy the compute endpoint base URL
  • Enable auto-top up from the agent wallet
  • Select the preferred chain used to fund top-ups
  • Set the top-up amount and low balance threshold

For agents created and run in the Virtuals Console, these compute credentials and endpoints are used automatically by the Console runtime. For external agents, copy the generated API key and endpoint into your own agent runtime or application.

Auto-Top Up

Auto-top up keeps compute available by replenishing the agent's compute balance from its agent wallet.

The preferred chain determines which chain's agent wallet funds the top-up. When the compute balance falls below the low balance threshold, the configured top-up amount is added automatically.

  • Auto Top Up Amount — the amount added when the compute balance is low
  • Low Balance Threshold — the balance level that triggers the top-up

Use conservative values for autonomous agents so compute stays available without authorizing larger wallet transfers than needed.

Using the Endpoint

Base URL

https://compute.virtuals.io/v1

Set your API key in VIRTUALS_API_KEY, then call the OpenAI-compatible chat completions endpoint directly.

OpenAI Chat Completions Format

Use /chat/completions when your agent or SDK already sends OpenAI-style chat messages.

curl
curl https://compute.virtuals.io/v1/chat/completions \
  -H "Authorization: Bearer $VIRTUALS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k2-0905",
    "messages": [
      { "role": "user", "content": "Hello from my agent" }
    ]
  }'

Codex Setup

Use the setup skill first. It installs the local Codex skill package, starts the Responses proxy, writes the Codex provider config, and verifies a fresh Codex request through Virtuals:

git clone https://github.com/Virtual-Protocol/acp-cli-demos.git
cd acp-cli-demos
scripts/install-local-skills.sh --mode symlink --target codex

Then open a fresh Codex CLI or Codex Desktop local thread and run:

$acp-builder-setup Configure Codex for Virtuals compute credits.

Under the hood, Codex custom providers use /v1/responses. Because the raw compute endpoint serves /v1/chat/completions, the skill configures the local proxy and provider block:

Proxy utility
cd acp-cli-demos/utilities/model-routing/codex-virtuals-proxy
cp .env.example .env
# edit .env and set VIRTUALS_API_KEY
npm start

The same Codex configuration is used by the CLI and Codex Desktop app local threads. Keep the proxy running while Codex is active. Do not paste real API keys into prompts; enter them in local files or environment variables when the skill asks.

Claude Setup

Use the setup skill first. It installs or checks the router, writes the maintained config, and verifies a real Claude Code request through Virtuals:

git clone https://github.com/Virtual-Protocol/acp-cli-demos.git
cd acp-cli-demos
scripts/install-local-skills.sh --mode symlink --target claude

Then open Claude Code and run:

/acp-builder-setup Configure Claude Code for Virtuals compute credits.

Under the hood, Claude Code terminal uses claude-code-router:

npm install -g @anthropic-ai/claude-code
npm install -g @musistudio/claude-code-router
 
mkdir -p "$HOME/.claude-code-router"
cp utilities/model-routing/claude-virtuals-router/config.example.json \
  "$HOME/.claude-code-router/config.json"
 
export VIRTUALS_API_KEY=...
ccr code

For Claude Desktop or Claude web, use the uploadable ZIP skills instead of the terminal router. The live checkout skill is intentionally terminal-only because it assumes local acp-cli, browser automation, card issuance, and checkout controls.

See the full setup matrix in acp-cli-demos/docs/agent-setup.md.