Agent Workflow — MCP + Skills

Quick summary

The AI agent uses MCP tools to execute tasks while Skills describe the safe order. A useful workflow is inspect → plan → user approval → apply → verify.

Confirm which MCP server the agent connects to and keep write approval, workspace trust, and evidence reporting explicit.

The intended way for an AI agent to drive the platform: MCP servers provide the tools; skills/*/SKILL.md files provide the playbooks.

Protect a module (local)

// MCP client config — yp-mcp
{ "mcpServers": { "yp": {
  "command": "node",
  "args": ["/abs/nougat-devpay/tools/devpay-mcp/server.mjs",
           "--workspace", "/abs/project"]
} } }

Flow (mirrors the protect_project prompt and yetdapay-protect skill):

1. yetdapay_environment_doctor — fix any FAIL.
2. yetdapay_project_analyze — report findings.
3. yetdapay_protection_plan — report decisions + plan_sha256 etc.
4. yetdapay_protection_review — show the user the summary and the shas; ask for explicit approval.
5. yetdapay_protection_apply with approved:true — refused without it, before the CLI spawns.
6. yetdapay_protection_verify — read .devpay/evidence/; report verdicts literally.

Install donation/feedback (remote)

// MCP client config — gateway
{ "mcpServers": { "yetdapay": {
  "url": "https://yetdapay-mcp-staging.<subdomain>.workers.dev/mcp",
  "headers": { "Authorization": "Bearer <seller-token>" }
} } }

Flow (mirrors setup_donation_and_feedback + yetdapay-install): yetdapay_store_list → yetdapay_integration_status / yetdapay_install_manifest_get → yetdapay_donation_configure + yetdapay_feedback_configure → yetdapay_integration_assets_get → apply the snippet/component → yetdapay_integration_validate → report.

Administer licenses (remote)

yetdapay_license_list / yetdapay_license_get / yetdapay_license_bindings / yetdapay_license_usage / yetdapay_usage_counters / yetdapay_security_events for inspection; yetdapay_license_verify for public key checks — the yetdapay-license skill encodes the seller-vs-public split.

Rules the agent must keep

• Only public_id in visitor-facing code — never credentials.
• Read-only tools before writes; writes are exactly yetdapay_donation_configure / yetdapay_feedback_configure / yetdapay_protection_apply — the last is approval-gated.
• Never retry around a refusal; PARTIALLY_VERIFIED is not success; AGENT/MANUAL_REVIEW decisions produce no code.