An autonomous content marketing team. Talk to your agents on Telegram. Always on. Auto-improving.
BrandCrew is an open-source multi-agent system that runs your entire LinkedIn content pipeline — from niche research to post design — with AI agents you direct through Telegram. Every week, the agents research, strategize, write, score, design, and deliver. Every cycle, they get smarter from your feedback.
7 full-time agents run on a VPS via cron. Each has a defined role, reads specific files, writes to specific outputs, and never talks to another agent directly. The filesystem is the communication layer.
Sunday Research → Direction → 10 Ideas → You pick (Telegram)
Mon–Fri Write → Score (35/50 pass) → You approve → Design → Deliver
Every cycle Marketing Director reads all feedback, adjusts strategy
You make 3 decisions per week via Telegram:
- Pick ideas — 10 ideas arrive, you multi-select your favorites
- Approve drafts — read the post, approve or send notes
- Review designs — see the image, approve or request changes
Everything else is automated.
| Code | Agent | Role |
|---|---|---|
| FT.01 | Niche Content Researcher | Scans your niche weekly for trending topics |
| FT.02 | Marketing Director | Synthesizes all data into weekly strategic direction |
| FT.03 | Content Strategist | Generates 10 ideas from research for your selection |
| FT.04 | Content Creator | Writes posts in your voice from approved ideas |
| FT.05 | Content Editor | Scores drafts on 5 dimensions — 35/50 to pass |
| FT.06 | Social Media Designer | Selects templates and renders branded visuals via Playwright |
| FT.07 | Content Repurposer | Adapts posts for X, newsletters, blog |
Plus 4 part-time research agents for deep niche intelligence, and 1 interactive agent for extracting your expertise through guided interviews.
This is not a "generate and forget" system. Three feedback loops run continuously: Built on an auto-research framework — agents continuously gather niche intelligence, analyze what's working, and feed insights back into the pipeline before any content is created.
- Content quality loop — Every draft is scored. Scores feed back into the Marketing Director's weekly analysis. Low-performing patterns get flagged. High-performing patterns get amplified.
- Design quality loop — You review every image. Feedback goes to the Designer's rules file. The system learns your visual preferences over time.
- Strategic direction loop — You can send strategic notes anytime via
/marketing_directoron Telegram. The Director reads these alongside performance data to adjust the entire pipeline's direction.
The agents don't have memory. They have files — shared knowledge files that get updated based on feedback, read by the agents that need them, and version-controlled in this repo.
| Component | What | Cost |
|---|---|---|
| AI | Claude (via Claude Code) | API usage |
| Database | Supabase (Postgres) | Free tier |
| Server | Hetzner VPS (Ubuntu) | ~$5/mo |
| Scheduling | Cron | Free |
| Communication | Telegram Bot | Free |
| Design | HTML templates + Playwright | $0 |
| Code | This repo | Open source |
No proprietary frameworks. No LangChain. No vector databases. Agents are SKILL.md files executed by Claude Code via bash scripts. The simplest architecture that works.
brandcrew/
├── agents/
│ ├── research/ # FT.01 — Niche researcher
│ ├── marketing_director/ # FT.02 — Weekly strategist
│ ├── ideation/ # FT.03 — Idea generator
│ ├── writer/ # FT.04 — Post creator
│ ├── quality/ # FT.05 — Scoring editor
│ ├── designer/ # FT.06 — Visual designer
│ ├── repurposer/ # FT.07 — Multi-platform adapter
│ └── ip_extractor/ # IA.01 — Expertise interviewer
├── shared/ # Shared knowledge files (agent memory)
│ ├── voice.md # Your voice profile
│ ├── performance-patterns.md # Weekly direction
│ ├── brand-guidelines-skill.md # Design rules
│ └── quality-standards.md # Anti-slop filter
├── templates/ # 6 HTML post templates
├── telegram/ # Bot with 12 commands
├── scripts/ # Design renderer (Playwright)
├── supabase/ # Schema (5 tables)
└── docs/ # Research reports & blueprints
Every agent folder contains a SKILL.md (instructions) and run.sh (cron runner).
BrandCrew is configured for a specific niche through a few files:
shared/niche.md— audience, keywords, competitive landscapeshared/voice.md— your writing voice and perspectiveshared/boundaries.md— topics that are off-limits
Swap these files and the entire system adapts to a different B2B industry.
