PGChat 🐘

Chat with your PostgreSQL database using natural language. Powered by local LLMs via Ollama or Claude by Anthropic.

Features

• Natural language to SQL — just ask questions about your data
• Beautiful terminal UI built with Rich
• Persistent named sessions with conversation memory
• Schema cached at startup — no redundant fetches mid-conversation
• Single SQL execution gateway with read-only safety enforcement
• Supports Ollama (local, private) and Anthropic Claude as backends
• pip installable — one command to get started

Quick Start

pip install pgchat
pgchat

On first run, a setup wizard asks for your DB connection and preferred model.

Installation

From PyPI

pip install pgchat

From source

git clone https://github.com/icon-gaurav/pgchat.git
cd pgchat
pip install -e .

Backends

Ollama (default — local and private)

Install Ollama from ollama.com, pull a model, then run pgchat.

Recommended models: qwen2.5:7b, llama3.1:8b, mistral:7b

ollama pull qwen2.5:7b
pgchat --model qwen2.5:7b

Anthropic Claude

export ANTHROPIC_API_KEY=your_key
pgchat --backend anthropic --model claude-haiku-3

Or set in your .env:

PGCHAT_BACKEND=anthropic
PGCHAT_MODEL=claude-haiku-3
ANTHROPIC_API_KEY=sk-ant-...

CLI Reference

pgchat                              Start interactive chat (default session)
pgchat --config                     Re-run connection setup wizard
pgchat --db-url <url>               Connect via DATABASE_URL
pgchat --model <name>               Override LLM model
pgchat --backend ollama|anthropic   Choose LLM backend
pgchat --session <name>             Open a named session
pgchat --show-tool-calls            Show raw tool call/response blocks
pgchat --version                    Show version

pgchat sessions list                List all saved sessions
pgchat sessions delete <name>       Delete a session
pgchat sessions export <name>       Export session as markdown

In-Chat Commands

Command	Description
/sessions	List all saved sessions
/new	Start a fresh session
/resume <name>	Switch to a different session
/clear	Clear current session history
/history	Show last 10 turns
/export	Export session as markdown
/refresh-schema	Re-fetch schema from the database
/help	Show available commands
exit	Quit PGChat

How Memory Works

Sessions are saved as JSON files in sessions/.

Schema is fetched once at startup and injected as context — the agent knows your full database structure before you ask your first question.

When a session exceeds 20 turns, older messages are summarized automatically to keep context tight.

Safety

All SQL runs through a single gateway function with a read-only safety check. Only SELECT, WITH, SHOW, and EXPLAIN queries are allowed.

cursor.execute() exists exactly once in the codebase — inside db.py.

Contributing

PRs welcome. Before submitting, verify the single-gateway constraint:

grep -rn "cursor\.execute\|conn\.execute" . --include="*.py"

Must return exactly one result: inside db.py.

Development Setup

git clone https://github.com/icon-gaurav/pgchat.git
cd pgchat
python -m venv .venv
.venv\Scripts\activate    # Windows
# source .venv/bin/activate  # Linux/Mac
pip install -e .
pytest

License

MIT — see LICENSE for details.

---

Built with ❤️ by Gaurav