메뉴
Set up ACP builder workflows for Codex, Claude Code, and Claude Desktop. Use when installing ACP skills, configuring Codex or Claude to use Virtuals-hosted models, or checking builder environment readiness.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Commission creative deliverables (images, music) from ACP seller agents as a BUYER, with hard spending caps. Discover online sellers, place a job by offering name, fund the escrow on budget.set, retrieve the deliverable URL, complete the job on Base. Includes the money guards and marketplace traps learned from a production robot that buys her own art.
Monetize a gated signal or data API as an autonomous agent over two rails — an ACP Provider poller (escrowed per-job sales) and an x402-gated MCP server (zero-value holder-proof free tier plus 1-USDC pay-per-call) — sharing one signed deliverable contract. Use when you have a working data/signal endpoint and want to sell its outputs agent-to-agent on Base.
Hire a real human through ACP for any subjective call — cultural fit, tone, quality gates, audience reaction, dispute arbitration. Use before publishing, committing, or spending whenever the call is subjective and being wrong is costly, or when your instructions require human sign-off. Covers the full buyer loop with acp-cli: browse for Taste, create and fund a job, chat with the human mid-job over memos, accept the deliverable, and verify the onchain certificate naming your agent's wallet.
Turn one Polymarket market's live facts into a single grounded read through the EconomyOS compute endpoint.
Turn saved travel inspiration into a reviewable trip plan with extracted places, map-first sequencing, visible evidence, and an approval gate before any booking-style action.
Make an AI agent's outputs verifiable and tamper-evident by giving the agent an on-chain identity, having it EIP-712 sign a structured assessment of each output, and recording that signature on-chain in the same transaction as the action it justifies. Use when a skeptic should be able to confirm the agent authored a specific output and that the record cannot be edited afterward.
| name | acp-builder-setup |
| description | Set up ACP builder workflows for Codex, Claude Code, and Claude Desktop. Use when installing ACP skills, configuring Codex or Claude to use Virtuals-hosted models, or checking builder environment readiness. |
Use this skill to help a builder get ready to run ACP demos and skills. There are two separate jobs, and they have very different stakes:
Do flow 1 for everyone. Only do flow 2 if the user wants to use Virtuals' free inference credits instead of their own Anthropic/OpenAI account.
First identify the target surface — it decides which commands apply:
acp-cli, ~/.agents/skills, and ~/.codex/config.toml.acp-cli, ~/.claude/skills, and claude-code-router.~/.claude/skills or claude-code-router, so flow 2 (routing) does not apply — use the ZIP upload path instead.If the target surface is unclear, ask which tool they are setting up before changing files.
This installs the demo skills into the runtime. It only copies or links files — no global config is touched.
Use this repo as the source of truth. Prefer symlinks for local development and
copies for one-off installs. For exact commands, read references/setup-commands.md.
Do not assume repo-scoped .agents/skills or .claude/skills folders exist.
This repo keeps canonical skills directly under skills/, then installs or links
them into each local runtime.
Install these local-execution skills for Codex CLI/Desktop and Claude Code:
acp-builder-setupacp-paid-subscription-checkoutInstall these Claude Desktop upload packages:
packages/claude-desktop/acp-builder-setup.zippackages/claude-desktop/acp-paid-subscription-checkout.zipIn Claude Desktop, the combined checkout skill must use handoff or evidence-review mode unless the user has a separate local-tool bridge. Live checkout execution assumes local command execution, browser automation, and live payment controls.
Optional perk: point the agent at Virtuals-hosted models so inference is billed to
Virtuals credits instead of the user's own account. Skip this entirely on
Claude Desktop — it can't use ccr or the proxy.
The exact commands and the lifecycle diagrams live here:
references/setup-commands.md — copy-paste commands (bundled with this skill).docs/agent-setup.md — human walkthrough with diagrams, the three moving parts (config switcher / proxy / agent runtime), and recovery.Your job is to run that lifecycle safely. Three behaviors:
Explain the mental model first. Routing ON → the agent spends Virtuals credits, not the user's Anthropic/OpenAI account (even though it still looks like "their" Claude/Codex). Routing OFF → back on their own account. Fully reversible.
Confirm before mutating config. The lifecycle is make claude-on / make claude-off (Claude Code) and make codex-on / make codex-off (Codex), run from
the repo root with VIRTUALS_API_KEY set. These edit the user's global config
(~/.claude-code-router/config.json, ~/.codex/config.toml), so treat them like
provisioning: recommend the exact command for the detected surface, say it's
reversible, and wait for explicit authorization. make claude-check is read-only
— run it freely.
Always offer teardown. Whenever routing is on (or you just turned it on), tell
the user the one command back to their own account: make claude-off or make codex-off. If teardown fails, point them at the "Recover" section of
docs/agent-setup.md.
When local terminal tools are available:
VIRTUALS_API_KEY is set before starting model-routing utilities.acp-cli is installed before running ACP workflows.End with:
make claude-off / make codex-off).