A deterministic run / fix_first / kill verdict for every ad — before you spend on it, not after.
Spendict is an ad creative scoring API that judges every variant across seven dimensions — hook strength, angle, clarity, audience fit, platform fit, CTA, and compliance — and returns a server-verified verdict your agent can branch on. Connect via MCP (Claude, Cursor, ChatGPT, n8n) or plain REST. 100 free verdicts, no credit card required.
MCP · CLI · Skill · REST · 100 free verdicts · about a penny a call after
“Cut your ad-spend leak in 30 days.”
“Unlock the AI advantage today ✨”
“Founders: your CAC is lying to you.”
“The #1 secret marketers don't want you to know”
“We saved a DTC brand $42k in Q3.”
“Meta CPMs up 34%. Here's what still works.”
“Try our platform. It's really good.”
“Your PMax is fighting your Search. Fix it.”
“Stop A/B testing headlines. Start testing offers.”
“Scale to $1M/mo with one weird trick”
the whole batch, no cherry-picking — 3 run · 3 fix_first · 4 killed before a cent of spend
Drop in the Skill and your agent gates every ad automatically — or wire it up over MCP, the CLI, or plain REST. Same four tools, same verdict, same quota.
Same SKILL.md works in Claude Code, Cursor, Codex, and Gemini.
npx skills add spendict/skills
The skill calls Spendict over MCP or the CLI — connect once (see the MCP or CLI tab).
Your agent now gets a run / fix_first / kill on every ad before it recommends launching.
Four tools, one deterministic verdict each. The budget only moves on ads that earned it.
Most ad teams discover a bad creative after three days of spend and a CPM spike. Spendict’s ad creative scoring API moves that judgment to the moment the ad is written — one call returns a run, fix_first, or kill verdict with the single predicted failure mode attached. At roughly one cent per call, scoring 100 variants costs about a dollar — less than a single wasted click.
Explore pricing →Scores hook, angle, clarity, audience fit, platform fit, CTA, and compliance — and names the single predicted failure mode before a cent is committed.
Budget allocation, audience setup, bid strategy, measurement — checked against each platform's actual rulebook, before the structure fragments your spend.
Feed it live metrics and it separates creative fatigue from structural problems — instead of guessing which lever to pull.
Give it the product and the goal; it returns a structure-validated targeting strategy your agent can build from directly.
The judgment inside comes from performance marketers with 10+ years in paid social — they directed the model's development and calibrated every verdict against real ads.
Six AI-generated ad hooks, scored by the same engine that grades yours — each with the predicted failure mode attached, not just a number.
Every verdict includes a breakdown across seven creative dimensions: hook strength, angle specificity, clarity, audience resonance, platform fit, CTA, and compliance risk. Scores are deterministic — the same ad always returns the same verdict, making your QA pipeline auditable and reproducible across runs.
Try it on your own ad →“Unlock the AI advantage today ✨”
Zero specificity, generic AI puffery
“I switched from boosting posts to this and finally stopped guessing.”
Named pain plus concrete outcome
“Stop guessing. Start scaling your ad spend with confidence.”
Vague CTA, no proof point
“The future of marketing is here.”
No product, no audience, no hook
“3 signs your Meta campaign is bleeding budget (and how to fix #2 today)”
Specific plus curiosity plus a real fix
“Transform your business with our revolutionary platform.”
Buzzword salad, zero mechanism
How it works
POST the ad copy, target platform, and product context to a single endpoint — or let your MCP-connected agent call the tool directly. Text-only works; include a creative URL and the visual is assessed too. The payload is the same across MCP, CLI, and REST.
Spendict scores the creative across seven dimensions — hook strength, angle specificity, clarity, audience resonance, platform fit, CTA, and compliance risk — then the server computes a deterministic run, fix_first, or kill verdict. The model proposes; the server decides. The same ad always returns the same score.
Your agent branches on the verdict: launch the ad, revise the weak dimension, or drop it entirely. Each response includes the single predicted failure mode so the fix is targeted, not a guess. Over time, your reconciliation loop feeds real performance back into the system so the gate gets sharper.
Use cases
If your agent generates ad variants automatically — through Claude, GPT, or a custom pipeline — Spendict is the quality gate between generation and launch. The agent writes the ads; the ad creative scoring API decides which ones deserve budget. No human reviewer bottleneck, no bad ads reaching the auction.
A senior buyer can spot a weak hook in seconds — but not when there are 200 variants in the queue. Spendict encodes that judgment into a single API call: the same seven dimensions a human expert checks, scored consistently and instantly. Use it to triage the batch so human attention goes where it matters.
Build a workflow that generates creative on a schedule, scores each variant through the Spendict API, and only pushes the winners to your ad platform. The REST endpoint accepts a single POST with ad copy, platform, and product context — easy to wire into any automation tool that can make HTTP requests.
Each client has different compliance needs, audience expectations, and platform norms. Spendict checks every ad against platform-specific conventions and flags compliance risks before the creative reaches the client for approval — reducing revision rounds and protecting the agency from launching an ad that violates platform policy.
Four guardrails baked into every call — not settings you have to remember to turn on.
LLMs are persuasive — a confident model will explain why a bad ad is actually fine. Spendict removes that failure mode: the server, not the model, decides the verdict. Quota is checked before inference so a maxed-out key never triggers a paid model call, and every failed call is refunded automatically.
Streamable HTTP — works in Claude, Cursor, Windsurf, n8n out of the box.
Same verdicts, same key — one-for-one REST endpoints for any other stack.
Every key is hashed server-side; bearer-token auth on every call.
A maxed-out key never triggers a model call — and failed calls are refunded.
Real performance reconciles against the tool's own earlier pre-flight call.
Marketers with 10+ years in paid social directed the model's development — real ads, not scraped trend data.