Docs · Guides
Give your AI agent an economic decision step
A paste-in instruction, four open skills and the guardrails that keep an agent's spending in your hands.
Language models research, write and operate. What they cannot do on their own is choose and size across thousands of deals against your history. P34 is that step: your agent prepares the request, P34 decides, a person approves.
Paste this into your coding agent
For Claude Code, Codex or your IDE agent, working in your own repository:
Add a P34 decision step to this project. Read https://api.hyperc.com/v1/ and https://hyperc.com/docs.html first, and follow the field rules in https://github.com/hyperc-ai/P34-API-DOCS. Map our data into Menus (current options plus history, including the options we did not take) and Sales (what happened to the ones we took), and write a business description of our unit economics. Validate with POST /validate (no key) until there are no errors. Read the key from the P34_API_KEY environment variable; never write it into code, logs or URLs. Submit with mock: true first, poll GET /result/{session_id} until done or failed, and save the response. Show the qty > 0 rows as a plan for human review. Do not place orders.
Give it the skills
Four open skills carry the judgement calls the reference cannot state as rules, in the order a request goes through them: prepare inputs, business economics, submit and monitor, interpret results. They use the portable Agent Plugins layout; point your agent at the four SKILL.md files, and pin a commit so skills and docs move together.
Agents browsing the web should read hyperc.com/llms.txt first.
Guardrails
- Key: from an environment variable only — never in code, logs, prompts or URLs. One key per agent, so you can revoke it alone.
- Spend: top up only what you are willing to let the agent spend; the API refuses new fits once the balance is spent.
- Free first:
/validateand an input test before every actual fit. - Stop at the plan: the agent shows
qty > 0rows for review. P34 places no orders, and neither should the agent until you decide it may. - Feedback is input: a failed fit's
feedbackis written to be acted on — let the agent use it for the next iteration.