LinkedIn Agent

A multi-agent system that researches and drafts LinkedIn posts for company pages designed to work for any brand.

What it does

• Takes a topic as input and classifies it by domain, content type, and angle (orchestrator agent)
• Researches the topic via web search, scrapes real articles, and stores findings as embeddings in ChromaDB (research agent)
• Finds real-world examples that support the topic (examples agent)
• Flags weaknesses in the research and examples (critic agent)
• Synthesises everything into a structured content brief (brief agent)
• Generates 5 hooks across different formats — bold claim, stat, confession, question, one-liner (hook agent)
• Writes 2 full post drafts in different tones — punchy and narrative (draft agent)
• Scores both drafts on originality, promise fulfilment, and shareability (writing critic agent)

User picks the best hook and draft, edits it, and publishes.

Tech stack

• Python 3.11 — core language
• Ollama + gemma2:2b — local LLM, free, no API cost
• Pydantic — structured output validation between agents
• Serper API — web search (2,500 free searches/month)
• BeautifulSoup + requests — web scraping
• sentence-transformers (all-MiniLM-L6-v2) — text embeddings
• ChromaDB — persistent vector store for RAG
• FastAPI — REST API exposing the full pipeline

Setup

# 1. Clone the repo
git clone https://github.com/yourusername/linkedin-agent.git
cd linkedin-agent

# 2. Create and activate virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. Install Ollama and pull the model
# Download Ollama from ollama.com, then:
ollama pull gemma2:2b

# 5. Create your .env file
echo "SERPER_API_KEY=your_key_here" > .env
# Get a free key at serper.dev — 2,500 searches/month free

# 6. Create the data directory for ChromaDB
mkdir -p data

Running

CLI — runs the full pipeline in your terminal:

python main.py

API — starts the FastAPI server:

./venv/bin/fastapi dev api.py

Then go to http://127.0.0.1:8000/docs to use the interactive API docs.

Single POST endpoint:

POST /generate Body: { "topic": "Your Topic" } Returns: { "hooks": [...], "drafts": [...], "scores": {...} }

Project structure

linkedin-agent/ ├── agents/ │ ├── orchestrator.py # classifies topic into domain, content type, angle │ ├── research.py # web search, scraping, embedding, RAG retrieval │ ├── examples.py # finds real-world examples supporting the topic │ ├── critic.py # flags weaknesses in research and examples │ ├── brief.py # synthesises everything into a content brief │ ├── hook.py # generates 5 hooks across different formats │ ├── draft.py # writes 2 full post drafts in different tones │ └── writing_critic.py # scores drafts on originality, fulfilment, shareability ├── core/ │ ├── llm.py # Ollama wrapper — single call_llm function │ ├── models.py # Pydantic models for structured agent outputs │ └── utils.py # shared utilities — clean_topic, strip_json_fences ├── data/ │ └── chroma/ # persistent ChromaDB storage (gitignored) ├── main.py # CLI entry point — runs the full pipeline ├── api.py # FastAPI server — POST /generate endpoint ├── requirements.txt # all dependencies └── .env # SERPER_API_KEY (gitignored)

Status

Built and working:

• All 8 agents — orchestrator through writing critic
• Full pipeline via CLI (main.py) and REST API (api.py)
• Persistent ChromaDB — research chunks survive between runs
• JSON fence stripping and smart quote handling for gemma2:2b quirks

Next:

• Streamlit UI — frontend that calls the API and displays results
• Refinement loop — critic feeds back into research for better output
• Multi-tenant config — plug in any company's brand voice and categories