Successful marketing isn't about more API calls — it's about understanding what an image actually shows, how an email program is actually performing, and how the experts pull those pieces together. Agentic Tools is the toolkit that hands all three to your agent.
It can read a CDN URL. It can't tell you whether that hero shot says "weekend" or "wedding."
Eight weeks of sends, tone shifts, cadence drift, AI-readiness gaps — invisible without an evaluation layer.
Without an opinionated playbook, every agent run reinvents what "good" looks like — and lands somewhere safe but generic.
A digital asset manager built for agents that need to see. Search returns matched images as inline pixels alongside the URLs — so the model actually looks at the assets before it picks one, generates a derivative, or drops one into an email.
That sounds small until you watch a generic agent confidently pick the wrong photo because the filename had "hero" in it. SAM closes the gap: 46 tools across images, AI derivatives, products, and bulk jobs — all of them aware that the model has eyes.
{
"mcpServers": {
"cordial-sam": {
"type": "http",
"url": "https://sam-mcp.cordial.com/mcp",
"auth": "oauth"
}
}
}// search_images returns matched images as inline base64 blocks
// — the model literally looks at them before picking one.
const matches = await mcp.call("search_images", {
query: "red running shoe on white background",
aspect: "1:1",
top_k: 8,
});
// → 8 actual images + URLs + metadata in one turn
// Now the agent picks the one that fits the brand,
// not the one whose *filename* sounds right.
const variants = await mcp.call("smart_resize_image", {
image_id: matches.images[0].id,
target_sizes: ["1080x1080", "1200x628"],
aspect_ratios: ["16:9"],
});The hardest context to give an agent is what your program is actually doing — tone, cadence, offer mix, AI-readiness — over the last eight weeks. Understand.email ingests every send, runs it through an LLM analysis pipeline, and exposes the resulting corpus through an MCP server.
Plug it in and your agent knows: where the program is weak, which sends pulled it down, and what an A+ grade would even look like — before it writes a single line of copy.
{
"mcpServers": {
"understand-email": {
"type": "http",
"url": "https://understand.email/mcp",
"auth": "oauth"
}
}
}→ tool_call: account_status
← { ready: true, scope: "brand.com", paid: false }
→ tool_call: get_domain_audit { root_domain: "brand.com" }
← {
weighted_overall: 6.8, grade: "B-",
priority_focus_areas: [
{ category: "C2 · AI Summary Readiness", median: 4.1 },
{ category: "C3 · Schema.org / Annotations", median: 2.0 },
{ category: "C5 · Signal-to-Noise Ratio", median: 5.4 }
],
window: "8 weeks, ending 2026-05-24"
}
→ tool_call: list_domain_audit_emails
{ root_domain: "brand.com", max_score: 5, limit: 5 }
← [ { email_id: "e_91…", overall: 3.4, subject: "Last chance!" }, … ]
Now the agent knows what this program is actually shipping, where it's
weakest, and which sends to fix first — context no generic LLM has.Pixels and program data only get you halfway. The other half is knowing how to use them — how an expert email marketer would sequence the work, what tradeoffs to make, which signals actually matter. That craft lives in our team's heads. The Agentic Cookbook moves it into your agent.
A public GitHub repo of Claude Skills, recipes, and playbooks
that bundle the prompts, references, and tool-use patterns
for one specific marketing job. Drop a skill into ~/.claude/skills/, or
read the recipe and adapt it to your own stack.
# Pull the cookbook
git clone https://github.com/CordialExperience/agentic-cookbook
cd agentic-cookbook
# Install a skill into Claude Code
cp -r skills/<skill-name> ~/.claude/skills/
# Skills bundle the prompts, references, and tool-use patterns
# that make an agent actually competent at one specific job.Every tool here speaks Model Context Protocol — the open standard. Bring the agent you already use, or wire one of your own. We don't ship the assistant; we ship the context that makes yours competent.
Each scenario chains the three layers — visual, program, playbook — into something a generic LLM couldn't do alone.
The agent pulls the audit, identifies the worst grades and what's dragging them, suggests new copy and imagery on-brand, and stages the rebuilds.
understand.email · get_domain_audit understand.email · list_domain_audit_emails (max_score: 5) sam · search_images (replacements that fit the brand) cookbook · /skills/campaign-rebuild From a CSV to a full enriched, image-rich catalog with lifestyle shots generated from the existing product photography — without a creative-ops sprint.
sam · start_import (catalog CSV, enrich_on_write: true) sam · enrich_all_products sam · generate_product_image (lifestyle variants) cookbook · /skills/product-launch A focused roadmap based on the actual sends in your program — what's missing, what to fix first, and how to prove progress send-over-send.
understand.email · subscribe a domain understand.email · get_domain_audit_history cookbook · /skills/ai-readiness-roadmap Subscribe their public marketing addresses, then run a portfolio-wide read on cadence, tone, offer mix, and AI-readiness — at a level no generic LLM can produce.
understand.email · list_domains (paid scope) understand.email · get_domain_analytics (per competitor) cookbook · /skills/competitive-readout The cookbook is open source. Understand.email is free to try — no Cordial account required. SAM MCP runs the moment you have a SAM tenant.