defog
A comprehensive Python toolkit for AI-powered data operations - from natural language SQL queries to multi-agent orchestration.
Features
- 🤖 Cross-provider LLM operations - Unified interface for OpenAI, Anthropic, Gemini, Grok (xAI), and Together AI
- 📊 SQL Agent - Convert natural language to SQL with automatic table filtering for large databases
- 🔍 Data extraction - Extract structured data from PDFs, images, HTML, text documents, and even images embedded in HTML
- 🛠️ Advanced AI tools - Code interpreter, web search, YouTube transcription, document citations
- 🎭 Agent orchestration - Hierarchical task delegation and multi-agent coordination
- 💾 Memory management - Automatic conversation compactification for long contexts
Installation
pip install --upgrade defog
Quick Start
1. LLM Chat (Cross-Provider)
from defog.llm.utils import chat_async
from defog.llm.llm_providers import LLMProvider
# Works with any provider
response = await chat_async(
provider=LLMProvider.ANTHROPIC, # or OPENAI, GEMINI
model="claude-sonnet-4-20250514",
messages=[{"role": "user", "content": "Hello!"}]
)
print(response.content)
OpenAI GPT‑5: Responses API controls
from defog.llm.utils import chat_async
from defog.llm.llm_providers import LLMProvider
response = await chat_async(
provider=LLMProvider.OPENAI,
model="gpt-5.1",
messages=[
{"role": "system", "content": "You are concise and helpful."},
{"role": "user", "content": "Summarize the benefits of unit tests."},
],
# Optional Responses API controls for GPT‑5.1
reasoning_effort="none", # none | low | medium | high
verbosity="low", # low | medium | high
)
print(response.content)
2. Natural Language to SQL
from defog.llm.sql import sql_answer_tool
from defog.llm.llm_providers import LLMProvider
# Ask questions in natural language
result = await sql_answer_tool(
question="What are the top 10 customers by total sales?",
db_type="postgres",
db_creds={
"host": "localhost",
"database": "mydb",
"user": "postgres",
"password": "password",
"port": 5432
},
model="claude-sonnet-4-20250514",
provider=LLMProvider.ANTHROPIC
)
print(f"SQL: {result['query']}")
print(f"Results: {result['results']}")
3. Extract Data from PDFs
from defog.llm import extract_pdf_data
# Extract structured data from any PDF
data = await extract_pdf_data(
pdf_url="https://example.com/financial_report.pdf",
focus_areas=["revenue", "financial metrics"]
)
for datapoint_name, extracted_data in data["data"].items():
print(f"{datapoint_name}: {extracted_data}")
4. Code Interpreter
from defog.llm.code_interp import code_interpreter_tool
from defog.llm.llm_providers import LLMProvider
# Execute Python code with AI assistance
result = await code_interpreter_tool(
question="Analyze this data and create a visualization",
csv_string="name,sales\nAlice,100\nBob,150",
model="gpt-4o",
provider=LLMProvider.OPENAI
)
print(result["code"]) # Generated Python code
print(result["output"]) # Execution results
5. Using MCP Servers with chat_async
from defog.llm.utils import chat_async
from defog.llm.llm_providers import LLMProvider
# Use MCP servers for dynamic tool integration
# Works with both local and remote MCP servers
response = await chat_async(
provider=LLMProvider.OPENAI,
model="gpt-4.1",
mcp_servers=["http://localhost:8000/mcp"], # Can be local or remote
messages=[
{"role": "user", "content": "How many users are in the first table?"}
]
)
# MCP tools are automatically converted to Python functions
# and made available to the LLM
print(response.content)
6. Anthropic Server-Side Tools and Programmatic Tool Calling
chat_async exposes Anthropic's first-party server-side tools (web_search,
web_fetch, code_execution, advisor) and the new programmatic tool
calling flow, where Claude writes Python in the code execution sandbox that
calls your local tools as await my_tool(...) — keeping intermediate
results in the sandbox so they never re-enter the model context.
from pydantic import BaseModel
from defog.llm.utils import chat_async
# Server-side web_search
response = await chat_async(
provider="anthropic",
model="claude-opus-4-6",
messages=[{"role": "user", "content": "What's the latest defog-python release?"}],
server_tools=["web_search"],
)
print(response.content)
print(response.server_tool_outputs) # raw web_search_tool_result blocks
print(response.server_tool_usage) # {"web_search_requests": 1, ...}
# Programmatic tool calling: Claude calls your tool from inside code execution
class QueryArgs(BaseModel):
sql: str
async def query_database(input: QueryArgs) -> list:
"""Run a SQL query and return rows as JSON."""
return [{"customer": "Acme", "revenue": 50_000}]
response = await chat_async(
provider="anthropic",
model="claude-opus-4-6",
messages=[{"role": "user", "content": "Who is the top customer by revenue?"}],
tools=[query_database],
server_tools=["code_execution"],
programmatic_tool_calling=True,
)
print(response.content)
print(response.container_id) # reuse via `container_id=` on a follow-up call
# Task budgets: advisory token cap across the full agentic loop
# (Claude Opus 4.7 only). Accepts an int (expands to {"type": "tokens",
# "total": N}) or the full dict with optional "remaining" for loops that
# compact history between requests. Minimum 20,000 tokens.
response = await chat_async(
provider="anthropic",
model="claude-opus-4-7",
messages=[{"role": "user", "content": "Audit this repo for security issues."}],
tools=[...],
task_budget=64000,
)
See docs/llm/anthropic-server-tools.md
for the full reference, including version overrides for Bedrock/Vertex,
container reuse, and the LLMResponse shape additions.
Documentation
📚 Full Documentation - Comprehensive guides and API reference
Quick Links
- LLM Utilities - Chat, function calling, structured output, memory management
- Database Operations - SQL generation, query execution, schema documentation
- Data Extraction - PDF, image, and HTML data extraction tools
- Agent Orchestration - Multi-agent coordination and task delegation
- API Reference - Complete API documentation
Environment Variables
# API Keys
export OPENAI_API_KEY="your-openai-key"
export ANTHROPIC_API_KEY="your-anthropic-key"
export GEMINI_API_KEY="your-gemini-key"
Advanced Use Cases
For advanced features like:
- Memory compactification for long conversations
- YouTube video transcription and summarization
- Multi-agent orchestration with shared context
- Database schema auto-documentation
- Model Context Protocol (MCP) support
See the full documentation.
Development
Testing and formatting
- Run tests:
python -m pytest tests - Format code:
ruff format - Update documentation when adding features
Releasing
Bump version in pyproject.toml in the PR that should ship. When the PR merges, the Publish to PyPI workflow (.github/workflows/publish.yml) builds the package with uv build and uploads it to PyPI through trusted publishing. Merges that leave the version unchanged, or whose version is already on PyPI, are skipped.
Using our MCP Server
- Run
defog serveonce to complete your setup, anddefog dbto update your database credentials - Add to your MCP Client
- Claude Code:
claude mcp add defog -- python3 -m defog.mcp_server. Or if you do not want to install the defog package globally or set up environment variables, runclaude mcp add dfg -- uv run --directory FULL_PATH_TO_VENV_DIRECTORY --env-file .env -m defog.mcp_server - Claude Desktop: add the config below
{ "mcpServers": { "defog": { "command": "python3", "args": ["-m", "defog.mcp_server"], "env": { "OPENAI_API_KEY": "YOUR_OPENAI_KEY", "ANTHROPIC_API_KEY": "YOUR_ANTHROPIC_KEY", "GEMINI_API_KEY": "YOUR_GEMINI_KEY", "DB_TYPE": "YOUR_DB_TYPE", "DB_HOST": "YOUR_DB_HOST", "DB_PORT": "YOUR_DB_PORT", "DB_USER": "YOUR_DB_USER", "DB_PASSWORD": "YOUR_DB_PASSWORD", "DB_NAME": "YOUR_DB_NAME" } } } }
- Claude Code:
Available MCP Tools and Resources
The Defog MCP server provides the following capabilities:
Tools (actions the AI can perform):
text_to_sql_tool- Execute natural language queries against your databaselist_database_schema- List all tables and their schemasyoutube_video_summary- Get transcript/summary of YouTube videos (requires Gemini API key)extract_pdf_data- Extract structured data from PDFsextract_html_data- Extract structured data from HTML pagesextract_text_data- Extract structured data from text files
Resources (read-only data the AI can access):
schema://tables- Get list of all tables in the databaseschema://table/{table_name}- Get detailed schema for a specific tablestats://table/{table_name}- Get statistics and metadata for a table (row count, column statistics)sample://table/{table_name}- Get sample data (10 rows) from a table
License
MIT License - see LICENSE file for details.
1 maintainer
Metadata
1 maintainer
Release files for defog 1.6.17
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| defog-1.6.17.tar.gz | 331.0 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| defog-1.6.17-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 585.7 kB
Release files / defog-1.6.17.tar.gz
| Download URL | defog-1.6.17.tar.gz |
|---|---|
| Size | 331.0 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e41b3d181fe183d48a6768bee2b6419c7fcefc2a59235d0a7b64314781b67c24
|
|
BLAKE2b-256 checksum How to use checksums |
1ec0763053dc251e5ddb7fa787d15e6eedbef11ab853c1809a089ab0fd596b7a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 15, 2026.
Transparency logRelease files / defog-1.6.17-py3-none-any.whl
| Download URL | defog-1.6.17-py3-none-any.whl |
|---|---|
| Size | 254.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
4d4b7ff00ad15cf60c110ff8f41798c8d4202450c9212763bd8ace93cb397637
|
|
BLAKE2b-256 checksum How to use checksums |
d887e5eaf729c2a9cf3d271ea41df7109f5124cf3d8c01e2366a065f9baf025c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 15, 2026.
Transparency log
