An autonomous AI marketing team, built on the Claude Agent SDK. Give it a product. Every day the team plans, creates and ships marketing work: campaigns, banners, social posts, short-form videos, App Store listings, ad plans, press kits. Everything lands as files you review and approve.
A CMO agent leads 11 specialist subagents. It reads the team's memory, decides the day's priorities, delegates work in parallel, sends everything to QA and writes the day's report.
| Agent | Owns |
|---|---|
| 🧭 CMO (orchestrator) | Daily plan, delegation, memory, final report |
| 🔎 market-researcher | Trends, competitors, audience insights, cultural moments (web search) |
| 📈 growth-strategist | Positioning, messaging, KPIs, channel and budget mix, ICE experiment backlog |
| 📱 aso-specialist | App Store and Google Play listings, keywords, screenshots, release notes |
| ✍️ copywriter | Ads, landing pages, email, push, in-app copy with A/B variants |
| 💬 social-media-manager | Platform-native posts for IG, TikTok, X, LinkedIn, Threads, Reddit and YouTube |
| 🎨 visual-designer | Banners, ad creatives, carousels and thumbnails as real PNG files |
| 🎬 video-producer | Short-form scripts and storyboards, rendered MP4s, shot lists |
| 💰 performance-marketer | Paid plans for Meta, Google App, Apple Search Ads, TikTok and Reddit, with bulk CSVs |
| 📰 pr-community-manager | Press releases, Product Hunt/HN/Reddit launches, creator outreach |
| 📊 data-analyst | Reads your metric exports, tracks KPIs, says what to scale or stop |
| 🛡 brand-guardian | Final QA for brand voice, claims, character limits and accessibility |
The team gets in-process MCP tools that produce actual assets:
render_banner: on-brand creatives in 19 preset sizes (Instagram, Stories, X, LinkedIn, display ads, YouTube thumbnails, OG images, Play feature graphic, App Store screenshots…) using your brand colors, font, logo and screenshots.render_video: vertical, square or landscape MP4 videos with kinetic typography (pop, rise, slide, type-on), screenshot scenes, a progress bar and an end-card CTA. It also writes a cover PNG and SRT captions.log_deliverable: adds each asset to a manifest with owner, channel, status and suggested publish time.summarize_metrics: profiles the CSV exports you drop intometrics/.
git clone https://github.com/pballada/mktagent && cd mktagent
pip install -e . # Python 3.10+
export ANTHROPIC_API_KEY=sk-ant-... # or log in once with the Claude Code CLI
cp examples/product.example.yaml my-product.yaml # describe your product
mktagent init my-product.yaml # creates workspaces/<product>/
mktagent kickoff workspaces/<product> # day 1: research, strategy, calendar, first content
mktagent daily workspaces/<product> # every day after thatOpen workspaces/<product>/index.html to see the review board with every banner, video and post,
or read days/<date>/README.md for the day's plan and your human to-do list.
| Command | What it does |
|---|---|
mktagent init brief.yaml |
Create a workspace from a product brief |
mktagent kickoff WS |
Research, strategy, 14-day calendar, ASO audit, brand starter pack, first video and first posts |
mktagent daily WS [--focus "TikTok"] |
Daily run with a weekly rhythm (Mon plan, Tue produce, Wed acquire, Thu PR/community, Fri retention…) |
mktagent campaign WS "Black Friday: 50% off annual" |
Full integrated campaign: plan, key visuals, 2 videos, copy, social calendar, paid plan, PR |
mktagent launch WS [--release 3.0] |
Release launch kit: What's New, trailer, carousel, press release, Product Hunt kit, T-14 to T+14 timeline |
mktagent ask WS "Write 10 TikTok hooks" |
Ad-hoc request; the CMO picks the right specialists |
mktagent gallery WS |
Rebuild the HTML review board |
mktagent team |
Show the roster |
Run flags: --model, --subagent-model (inherit / sonnet / haiku / full id), --effort
(low … max), --budget (hard USD cap per run), --max-turns, --date, --dry-run (prints the
prompts and spends nothing), --allow-bash, -v.
Each workspace has a persistent memory the team reads at the start of every run and updates at the end:
workspaces/pocketbrew/
├── product.yaml # your brief (edit anytime)
├── assets/ # logo, font, screenshots (copied from the brief)
├── metrics/ # drop CSV exports here: store, ads, social, GA4, RevenueCat…
├── memory/
│ ├── strategy.md # positioning, messaging house, channel mix, KPI targets
│ ├── calendar.md # rolling 14-day content & campaign calendar
│ ├── backlog.md # ICE-scored growth experiments
│ ├── learnings.md # what worked / didn't
│ └── log.md # what shipped each day
├── days/2026-11-02/
│ ├── README.md # plan, deliverables table, human to-do, tomorrow
│ ├── manifest.json # every deliverable with status
│ ├── qa-report.md # brand-guardian verdicts
│ ├── social/ banners/ video/ ads/ copy/ aso/ pr/ email/ research/
├── campaigns/black-friday/ …
└── index.html # review board
Feed it results (drop a CSV into metrics/). The data-analyst picks them up, and the strategy,
calendar and backlog change based on what performed.
GitHub Actions (included in .github/workflows/daily-marketing.yml):
- Fork this repo and create a workspace inside it:
mktagent init brief.yaml --dir marketing, then commitmarketing/. - Add the secret
ANTHROPIC_API_KEYand the variableMKTAGENT_WORKSPACE=marketing/<product>(optionallyMKTAGENT_DAILY_BUDGET, default $15). - The team runs every morning at 07:00 UTC and commits the day's deliverables. You can also trigger it manually with a mission and focus.
cron (local machine or server):
0 8 * * * cd /path/to/mktagent && .venv/bin/mktagent daily workspaces/pocketbrew --budget 15 >> mktagent.log 2>&1- Nothing is published automatically. The team produces drafts and a human to-do list. It never posts, sends emails, contacts people or spends ad budget.
- Agents can only read and write inside the workspace (a
PreToolUsehook enforces it). Shell access is off unless you pass--allow-bash. - Agents are told never to invent metrics, testimonials or contacts, and brand-guardian checks every deliverable for this.
- Use
--budgetto cap spend per run (enforced by the SDK). Use--subagent-model sonnetor--effort mediumfor cheaper daily runs.
| Env var | Default | |
|---|---|---|
ANTHROPIC_API_KEY |
– | Or an existing Claude Code login |
MKTAGENT_MODEL |
claude-opus-5-5 |
CMO model |
MKTAGENT_SUBAGENT_MODEL |
inherit |
Specialist model |
MKTAGENT_EFFORT |
high |
Reasoning effort |
Customize the team in src/mktagent/team.py (add a seo-specialist, an
influencer-manager…), the CMO's operating loop in src/mktagent/prompts.py
and the creative engine in src/mktagent/tools/.
pip install -e ".[dev]"
ruff check src tests && pytest # tests don't call the APIMIT
