# Peak Logic AI Engineering Toolkit > Pay-per-call AI engineering reviews for autonomous agents. No signup, no API > key, no subscription: each request is paid individually in USDC over the > x402 protocol (HTTP 402). Every endpoint returns structured JSON matching a > published schema, so responses can be consumed programmatically. Base URL: https://peakai.tools | | | |---|---| | Payment protocol | [x402](https://x402.gitbook.io/x402) (HTTP 402, `exact` scheme) | | Asset | USDC | | Network | Base mainnet (CAIP-2 `eip155:8453`) | | Receiving address | `0xD53254C04598844bfc7f543f8B6f3E15686E9b7C` | | Machine-readable schemas | `GET https://peakai.tools/catalog` (free) | | Rate limit | 300 requests/min per IP | ## How to pay 1. **Discover.** `GET /catalog` (free, no payment) returns every endpoint's price, input fields, and full output JSON Schema — enough to choose an endpoint and build a valid request without a human. 2. **Request.** POST your JSON body to the endpoint. Without payment you get `402 Payment Required`; the response carries the price, asset, network, and receiving address, plus the endpoint's expected request shape. 3. **Sign and retry.** Sign a USDC transfer authorization for the quoted amount and retry the same request with the signature in the `X-PAYMENT` header. An x402 client library does steps 2-3 automatically: ```js // npm install @x402/fetch @x402/evm viem import { wrapFetchWithPayment, x402Client } from "@x402/fetch"; import { registerExactEvmScheme } from "@x402/evm/exact/client"; import { privateKeyToAccount } from "viem/accounts"; const client = new x402Client(); registerExactEvmScheme(client, { signer: privateKeyToAccount(process.env.PRIVATE_KEY) }); const pay = wrapFetchWithPayment(fetch, client); const res = await pay("https://peakai.tools/scan-injection", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ content: "" }), }); const review = await res.json(); ``` The wallet only needs USDC on Base — no ETH for gas. x402's `exact` scheme uses gasless transfer authorizations; the facilitator submits the on-chain transaction. ### Payment guarantees - **You are only charged for a review you receive.** Settlement happens after a successful response; any error status (validation failure, model refusal, truncation, timeout, server error) cancels it. Nothing is transferred. - **Invalid requests are rejected before payment is taken.** Bodies are validated against the endpoint schema in a pre-payment hook; a malformed request returns the specific field errors, free of charge. - Every response includes `meta.price` (what the call cost) and the model that produced it. ## Endpoints Prices below are current for this deployment. All endpoints accept `Content-Type: application/json` and return structured JSON; the exact output schema for each is in `GET /catalog` under `output_schema`. ### POST /review-rag — $0.40 **RAG pipeline review.** Find the retrieval failures your eval set is not catching. Fields: - `description` (string, **required**) — What the RAG system does, the corpus, and the query patterns it serves. - `code` (string, optional) — Ingestion, chunking, and retrieval code. - `config` (string, optional) — Chunk sizes, embedding model, index params, top_k, reranker settings. - `known_problems` (string, optional) — Failures you have already observed, so the review can focus. Example request body: ```json { "description": "Support-ticket search over 40k Zendesk tickets. Users ask natural-language questions; we return an answer plus 3 cited tickets. Recall feels fine but answers cite the wrong ticket ~20% of the time.", "config": "chunk_size=1000 chars, overlap=0, embeddings=text-embedding-3-small, index=pgvector ivfflat lists=100, top_k=3, no reranker", "known_problems": "Multi-turn questions lose context. Acronym queries return nothing." } ``` ### POST /check-prompt — $0.25 **Prompt review and rewrite.** Get the failure modes in your prompt, plus a rewritten version. Fields: - `prompt` (string, **required**) — The system or user prompt to review. - `goal` (string, optional) — What the prompt is supposed to achieve, and for whom. - `model` (string, optional) — Target model, e.g. claude-opus-5. Changes the advice. - `observed_failures` (string, optional) — How it currently misbehaves. Example request body: ```json { "prompt": "You are a helpful assistant. CRITICAL: You MUST always use the search tool. Never make anything up. Answer the user's question about our product docs. Be concise but thorough and detailed.", "goal": "Answer product questions from docs, citing sources, in a support widget.", "model": "claude-opus-5", "observed_failures": "Searches even for 'hi'. Answers are 400 words when 40 would do." } ``` ### POST /check-architecture — $0.60 **System architecture review.** Where this design breaks, at what scale, and what to fix first. Fields: - `description` (string, **required**) — The system: components, data flow, storage, and what it is for. - `constraints` (string, optional) — Team size, budget, latency SLOs, compliance, existing stack. - `scale` (string, optional) — Current and expected traffic, data volume, growth rate. - `diagram` (string, optional) — Mermaid, ASCII, or a textual component list. Example request body: ```json { "description": "Next.js frontend on Vercel -> FastAPI on Render -> Postgres. A worker polls a jobs table every 5s to run LLM enrichment and writes results back. Redis caches responses for 10 minutes.", "constraints": "Two engineers, no dedicated ops. Must stay under $2k/mo.", "scale": "300 req/min peak, 50k enrichment jobs/day, growing 20% monthly." } ``` ### POST /review-linkedin — $0.20 **LinkedIn post review.** Honest read on whether the post earns attention, plus a rewrite. Fields: - `post` (string, **required**) — The draft post. - `audience` (string, optional) — Who you want to reach. - `goal` (string, optional) — What a successful post does: inbound leads, hiring, credibility. Example request body: ```json { "post": "I'm excited to announce that we've been working hard on something special. After months of effort, our team has launched a new AI-powered platform that leverages cutting-edge technology to revolutionize how businesses operate. Thoughts?", "audience": "Engineering leaders at Series A-C startups", "goal": "Inbound demo requests" } ``` ### POST /estimate-cost — $0.20 **LLM cost estimate.** Real arithmetic on your token spend, plus the levers that actually cut it. Fields: - `description` (string, **required**) — What the workload does, so the advice fits the use case. - `model` (string, optional) — Defaults to claude-opus-5. Known: claude-fable-5, claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5. - `requests_per_month` (integer, **required**) — Expected monthly request volume. - `avg_input_tokens` (integer, **required**) — Average prompt tokens per request. - `avg_output_tokens` (integer, **required**) — Average completion tokens per request, including thinking. - `cached_input_tokens` (integer, optional) — Size of the cacheable prefix. Default 0. - `cache_hit_rate` (number, optional) — 0-1 fraction of requests hitting that cached prefix. Default 0. - `custom_input_price_per_mtok` (number, optional) — Override for non-Anthropic or partner pricing. - `custom_output_price_per_mtok` (number, optional) — Override for non-Anthropic or partner pricing. Example request body: ```json { "description": "Support ticket classifier plus a drafted reply. Same 3k-token system prompt on every call.", "model": "claude-opus-5", "requests_per_month": 400000, "avg_input_tokens": 4200, "avg_output_tokens": 600, "cached_input_tokens": 3000, "cache_hit_rate": 0.85 } ``` ### POST /review-python — $0.40 **Python code review.** Bugs, concurrency hazards, and the failure you have not hit yet. Fields: - `code` (string, **required**) — The Python source to review. - `context` (string, optional) — How it is called, what runs it, expected inputs and scale. - `focus` (string, optional) — Narrow the review, e.g. 'async correctness only'. Example request body: ```json { "code": "import requests\n\ndef fetch_all(urls, cache={}):\n results = []\n for u in urls:\n if u not in cache:\n cache[u] = requests.get(u).json()\n results.append(cache[u])\n return results\n", "context": "Called from a FastAPI request handler, up to 200 urls per call." } ``` ### POST /review-docker — $0.25 **Dockerfile review.** Image size, cache misses, and the root user you forgot about. Fields: - `dockerfile` (string, **required**) — The Dockerfile contents. - `compose` (string, optional) — docker-compose.yml, if relevant. - `context` (string, optional) — Where it runs, base image constraints, build frequency. Example request body: ```json { "dockerfile": "FROM node:latest\nWORKDIR /app\nCOPY . .\nRUN npm install\nEXPOSE 3000\nCMD npm start\n", "context": "Deployed to Render. Rebuilt on every push to main." } ``` ### POST /review-kubernetes — $0.40 **Kubernetes manifest review.** Probes, limits, and the rollout that will drop traffic. Fields: - `manifest` (string, **required**) — Deployment, Service, Ingress, HPA — one or many YAML documents. - `context` (string, optional) — What the workload does, traffic shape, statefulness. - `cluster` (string, optional) — Managed provider, version, node sizes, existing policies. Example request body: ```json { "manifest": "apiVersion: apps/v1\nkind: Deployment\nmetadata:\n name: api\nspec:\n replicas: 1\n template:\n spec:\n containers:\n - name: api\n image: myorg/api:latest\n ports:\n - containerPort: 8080\n", "context": "Public HTTP API, ~200 rps, stateless." } ``` ### POST /review-vector-search — $0.40 **Vector search review.** Index parameters, recall cliffs, and filters applied in the wrong place. Fields: - `description` (string, **required**) — The corpus, the queries, and what 'correct' means. - `config` (string, optional) — Store, index type, dimensions, distance metric, build and search params. - `query_code` (string, optional) — How queries are built and executed. - `scale` (string, optional) — Vector count, growth, QPS, latency budget. Example request body: ```json { "description": "Semantic product search over 8M SKUs. Filter by category and in-stock before ranking. p95 budget 120ms.", "config": "pgvector 0.7, HNSW m=16 ef_construction=64, ef_search default, cosine, 1536 dims, filter applied in WHERE alongside ORDER BY embedding <=> query", "scale": "8M vectors, 400 QPS peak, growing 5%/mo" } ``` ### POST /review-agent — $0.45 **LLM agent review.** Tool surface, loop safety, and how this agent fails in production. Fields: - `description` (string, **required**) — What the agent does, who calls it, what it is allowed to touch. - `system_prompt` (string, optional) — The agent's system prompt. - `tools` (string, optional) — Tool definitions or schemas. - `observed_failures` (string, optional) — How it currently goes wrong. Example request body: ```json { "description": "Support agent that reads Zendesk tickets, searches our docs, and can issue refunds up to $500. Runs unattended on new tickets.", "tools": "search_docs(query), get_customer(email), issue_refund(order_id, amount_cents), close_ticket(id)", "observed_failures": "Occasionally closes tickets it did not resolve. Once refunded twice for the same order." } ``` ### POST /judge-output — $0.15 **LLM output judgment.** Independent grading of a model response against your rubric. Fields: - `task` (string, **required**) — The prompt or task the response was answering. - `response` (string, **required**) — The model output to judge. - `rubric` (string, optional) — Criteria to grade against, one per line. Without it, the judge derives criteria from the task. - `reference` (string, optional) — A known-good answer to compare against, if one exists. - `context` (string, optional) — What the score gates, e.g. 'regression eval, borderline should fail'. Example request body: ```json { "task": "Summarize this support ticket in 2 sentences, preserving the customer's requested refund amount.", "response": "The customer reports their order arrived damaged and is requesting a full refund of $89. They have attached photos and want a response within 2 business days.", "rubric": "1. Exactly 2 sentences.\n2. Refund amount stated and correct ($89).\n3. No invented details.", "context": "Gate for a prompt regression suite. Be strict: borderline fails." } ``` ### POST /scan-injection — $0.10 **Prompt injection scan.** Screen untrusted content before it reaches your agent's context. Fields: - `content` (string, **required**) — The untrusted content: webpage text, email, ticket, document, tool output. - `source` (string, optional) — Where it came from, e.g. 'scraped webpage', 'inbound email'. - `agent_context` (string, optional) — What the consuming agent can do — its tools and permissions — so severity reflects real blast radius. Example request body: ```json { "content": "Great product, works as described. ", "source": "product review scraped from a marketplace", "agent_context": "Summarizer agent with web_fetch and email tools, runs unattended." } ``` ### POST /diagnose-failure — $0.30 **Build/CI failure diagnosis.** Root cause from a failing log, and the fix — not just the error line. Fields: - `log` (string, **required**) — The failing output: CI log, build output, test run, stack trace. - `context` (string, optional) — Stack, CI system, when it last passed, whether it fails consistently. - `recent_changes` (string, optional) — The diff or commit list since the last green run. Example request body: ```json { "log": "npm ERR! peer dep missing: react@^18.0.0, required by @testing-library/react@14.1.2\n...\nTest suite failed to run\n Cannot find module 'react-dom/client' from 'src/setupTests.ts'", "context": "GitHub Actions, node 20. Passed yesterday; fails on every run since this morning. No lockfile in repo.", "recent_changes": "chore: bump testing-library packages" } ``` ## Free routes - `GET /` — human-readable documentation - `GET /catalog` — machine-readable service description (prices, input fields, output schemas) - `GET /agents.md` (alias `/llms.txt`) — this document - `GET /health` — liveness check