All 20 Criteria
P1-A Structured Data — 3/5
Homepage has schema.org WebSite and Organization markup with name, URL, description, and logo. No Offer, Product, or AggregateRating schema. Solid baseline, but pricing and product details are not schema-tagged.
P1-B Machine-Readable Pricing — 5/5
Every tool in /llm.txt includes: path, price in credits AND USD equivalent (e.g., "Price: 400 credits ($0.04)"), plus the master conversion rate "1 credit = $0.0001 USD (10,000 credits = $1.00)." An agent can compute exact cost for any tool call without human interpretation.
P1-C llms.txt / Agent Layer — 5/5
/llm.txt exists and is a complete agent-readable document: full tool catalog with descriptions, input/output schemas, per-tool pricing, CLI commands, MCP setup instructions, and REST API reference. Homepage also contains an identical hidden `<pre>` block for scraper access. Exceptional implementation.
P1-D API / MCP Availability — 5/5
Full OpenAPI 3.1.0 spec at /api/doc. MCP endpoint at https://agentpatch.ai/mcp (both Streamable HTTP and SSE transports). REST API with public endpoints (no auth required for discovery). Python CLI (`pip install agentpatch`). Claude Code skill and OpenClaw skill available for install.
P1-E Discoverability (GEO) — 3/5
robots.txt allows all public content, sitemap.xml present. Active blog with agent-relevant content. Published skills on agent platforms (Claude Code, OpenClaw) create discoverability within agent ecosystems. No explicit GEO/AI retrieval optimization beyond clean content.
P2-A Offer Completeness — 5/5
Every tool in /llm.txt is fully described: what it does, who uses it, input schema (field name, type, required/optional, description, valid values), output schema (field name, type, description), path, and exact price. All accessible from a single URL, parseable without human guidance.
P2-B Scope & Limits — 3/5
Topup limits stated (min $10, max $500). Failed call refunds documented. Some tools include explicit parameter caps (e.g., "max: 50" for result limits). However, rate limits (per-minute or per-day) are not documented in llm.txt or visible public docs — a gap for agents managing throughput.
P2-C Substitution Rules — 3/5
"Failed invocations (5xx, timeout) are fully refunded" is a clear and agent-legible fallback rule. Async job mechanism (poll /api/jobs/{job_id}) provides an explicit retry/poll pattern. No guidance on which tool to use if a specific tool category is unavailable.
P2-D Conditional Logic — 4/5
All pricing visible without authentication. Public endpoints (/api/tools, /api/search) require no API key. No hidden "contact sales" gates on standard tier. Minor gap: no published documentation of what happens if credit balance hits zero mid-task.
P2-E Semantic Precision — 4/5
Tool descriptions are specific and factual: "Search Amazon products. Returns product listings with prices, ratings, and ASINs." Input/output schemas define types explicitly. Output fields are named precisely. Homepage uses "context-optimized" without definition, but tool-level documentation is exemplary.
P3-A Verifiable Performance — 1/5
Self-reported claims only: "100% refund on failure," "hosted and maintained by AgentPatch." No third-party status page, no G2/Trustpilot reviews found. No uptime SLA published. Platform is early-stage (2026 launch), limiting historical track record.
P3-B Scoped Permissions — 2/5
One API key unlocks all 50+ tools simultaneously. Credit balance acts as an implicit spending cap. No per-tool permission scoping, no time-bounded keys, no action-specific restrictions. For an agent managing sensitive operations, this is an all-or-nothing model.
P3-C Audit Trail — 2/5
/api/jobs/{job_id} allows an agent to retrieve the status and output of any specific invocation by job ID — functional per-request traceability. No bulk audit log API or machine-accessible transaction history endpoint documented.
P3-D Behavioral Consistency — 1/5
No versioned terms, no changelog, no stated notice period for API changes or pricing updates. As an early-stage platform, this creates uncertainty for agents that need stable integration contracts.
P4-A Friction-Free Activation — 5/5
Signup → create API key in dashboard → set AGENTPATCH_API_KEY env var → call API. Homepage states "<30s Setup time." 10,000 free credits on signup. No human approval gate. CLI installable via pip. Agent can be operational in one shell session.
P4-B Agent Decision Signals — 4/5
Free "Hello World" tool (0 credits) lets an agent verify setup without cost. Every tool has explicit pricing enabling rational cost-benefit evaluation. Failed calls refunded, enabling safe experimentation. Missing: no programmatic signal for when usage patterns suggest subscription vs. pay-as-you-go optimization.
P5-A Integration Depth — 2/5
Claimed email addresses (@mail.agentpatch.ai) create identity lock-in: switching providers means losing the agent's email address and inbox history. Credit balance creates mild financial switching cost. No deep data sync, workflow history, or network effects.
P5-B Agent Memory Layer — 2/5
Email inbox (via claim-email-address + check-inbox tools) provides rudimentary persistent state for agent communications. No explicit memory or personalization API — each non-email tool call is stateless.
P5-C Programmatic Renewal — 4/5
POST /api/my/topup is explicitly documented in /llm.txt with min ($10) and max ($500) topup values. An agent can autonomously refill its credit balance without human intervention. This is a strong agent-native pattern.
P5-D Compounding Value — 2/5
"New APIs are added regularly" is stated in llm.txt. GET /api/tools returns a live catalog — an agent can programmatically discover newly added tools over time. No quantified signal of compounding value or integration-depth rewards for loyal agent users.
Rubric v1 (April 2026). Scores reflect the company's state on the audit date and may have improved since.