A Python framework for building agent applications with tools, RAG, persistent memory, guardrails, skills, file readers, browser automation, messaging platform integrations, and the Model Context Protocol. Switch providers without rewriting agent code.

---

Install

pip install definable

Or with uv:

uv pip install definable

Quick Start

from definable.agent import Agent
from definable.model.openai import OpenAIChat

agent = Agent(
  model=OpenAIChat(id="gpt-4o-mini"),
  instructions="You are a helpful assistant.",
)

output = agent.run("What is the capital of Japan?")
print(output.content)  # The capital of Japan is Tokyo.

Or use string model shorthand — no explicit import needed:

from definable.agent import Agent

agent = Agent(model="gpt-4o-mini", instructions="You are a helpful assistant.")
output = agent.run("What is the capital of Japan?")

Add Tools

from definable.agent import Agent
from definable.tool.decorator import tool


@tool
def get_weather(city: str) -> str:
  """Get current weather for a city."""
  return f"Sunny, 72°F in {city}"


agent = Agent(
  model="gpt-4o-mini",
  tools=[get_weather],
  instructions="Help users check the weather.",
)

output = agent.run("What's the weather in Tokyo?")

The agent calls tools automatically. No manual function routing.

Structured Output

from pydantic import BaseModel
from definable.agent import Agent
from definable.tool.decorator import tool


@tool
def get_weather(city: str) -> str:
  """Get current weather for a city."""
  return f"Sunny, 72°F in {city}"


class WeatherReport(BaseModel):
  city: str
  temperature: float
  conditions: str


agent = Agent(model="gpt-4o-mini", tools=[get_weather])

output = agent.run("Weather in Tokyo?", output_schema=WeatherReport)
print(output.content)  # JSON string matching WeatherReport schema

Pass any Pydantic model to output_schema and get validated, typed results back.

Streaming

from definable.agent import Agent

agent = Agent(model="gpt-4o-mini", instructions="You are a helpful assistant.")

for event in agent.run_stream("Write a haiku about Python."):
  if event.content:
    print(event.content, end="", flush=True)

run_stream() yields events as they arrive — content chunks, tool calls, and completion signals.

Multi-Turn Conversations

from definable.agent import Agent

agent = Agent(model="gpt-4o-mini", instructions="You are a helpful assistant.")

output1 = agent.run("My name is Alice.")
output2 = agent.run("What's my name?", messages=output1.messages)
print(output2.content)  # "Your name is Alice."

Pass messages from a previous run to continue the conversation.

Persistent Memory

from definable.agent import Agent
from definable.memory import Memory, SQLiteStore

agent = Agent(
  model="gpt-4o-mini",
  memory=Memory(store=SQLiteStore("memory.db")),
  instructions="You are a personal assistant.",
)

await agent.arun("My name is Alice and I prefer dark mode.", user_id="alice")
# Later, even in a new session...
await agent.arun("What's my name?", user_id="alice")  # Recalls "Alice"

Memory stores session history automatically and summarizes when messages exceed max_messages. For quick testing, use memory=True for an in-memory store. Three backends available: SQLite, file-based, and in-memory.

Knowledge Base (RAG)

from definable.agent import Agent
from definable.knowledge import Knowledge, Document
from definable.embedder import OpenAIEmbedder
from definable.vectordb import InMemoryVectorDB

kb = Knowledge(
  vector_db=InMemoryVectorDB(),
  embedder=OpenAIEmbedder(),
  top_k=3,
)
kb.add(Document(content="Company vacation policy: 20 days PTO per year."))

agent = Agent(
  model="gpt-4o-mini",
  instructions="You are an HR assistant.",
  knowledge=kb,
)

output = agent.run("How many vacation days do I get?")

The agent retrieves relevant documents before responding. Supports embedders (OpenAI, Voyage), vector DBs (in-memory, PostgreSQL, Qdrant, ChromaDB, MongoDB, Redis, Pinecone), rerankers (Cohere), and chunkers.

Note: Agent(knowledge=True) raises ValueError — unlike memory=True, knowledge requires explicit configuration with a vector DB.

Guardrails

from definable.agent import Agent
from definable.agent.guardrail import Guardrails, max_tokens, pii_filter, tool_blocklist
from definable.tool.decorator import tool


@tool
def get_weather(city: str) -> str:
  """Get current weather for a city."""
  return f"Sunny, 72°F in {city}"


agent = Agent(
  model="gpt-4o-mini",
  instructions="You are a support agent.",
  tools=[get_weather],
  guardrails=Guardrails(
    input=[max_tokens(500)],
    output=[pii_filter()],
    tool=[tool_blocklist({"dangerous_tool"})],
  ),
)

output = agent.run("What's the weather?")

Guardrails check, modify, or block content at input, output, and tool-call checkpoints. Built-ins include token limits, PII redaction, topic blocking, and regex filters. Compose rules with ALL, ANY, NOT, and when().

Skills

from definable.agent import Agent
from definable.skill import Calculator, WebSearch, DateTime

agent = Agent(
  model="gpt-4o-mini",
  skills=[Calculator(), WebSearch(), DateTime()],
  instructions="You are a helpful assistant.",
)

output = agent.run("What is 15% of 230?")

Skills bundle domain expertise (instructions) with tools. Built-in skills include Calculator, WebSearch, DateTime, HTTPRequests, JSONOperations, TextProcessing, Shell, FileOperations, and MacOS. Create custom skills by subclassing Skill.

MCP

from definable.agent import Agent
from definable.mcp import MCPConfig, MCPServerConfig, MCPToolkit

config = MCPConfig(
  servers=[
    MCPServerConfig(
      name="filesystem",
      command="npx",
      args=["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
    )
  ]
)

async with MCPToolkit(config=config) as toolkit:
  agent = Agent(model="gpt-4o-mini", toolkits=[toolkit])
  await agent.arun("List files in /tmp")

Connect to any MCP server. Use the same tools as Claude Desktop.

File Readers

from definable.agent import Agent
from definable.media import File

agent = Agent(
  model="gpt-4o-mini",
  readers=True,
  instructions="Summarize the uploaded document.",
)

output = agent.run("Summarize this.", files=[File(filepath="report.pdf")])

Pass readers=True to enable automatic parsing. Supports PDF, DOCX, PPTX, XLSX, ODS, RTF, HTML, images, and audio. AI-powered OCR available via Mistral, OpenAI, Anthropic, and Google providers.

Observability Dashboard

from definable.agent import Agent

agent = Agent(
  model="gpt-4o-mini",
  observability=True,  # enables live dashboard at /obs/
  instructions="You are a helpful assistant.",
)

agent.serve(enable_server=True)
# Open http://localhost:8000/obs/ in your browser

Live events (SSE), session history, run comparison, per-tool and per-model metrics — all in a single-page dashboard. Use ObservabilityConfig for fine-grained control (trace dir, buffer size, theme).

Deploy It

from definable.agent import Agent
from definable.agent.trigger import Webhook, Cron
from definable.agent.auth import APIKeyAuth

agent = Agent(model="gpt-4o-mini", instructions="You are a support agent.")

agent.on(Webhook(path="/support", method="POST"))
agent.on(Cron(schedule="0 9 * * *"))
agent.auth = APIKeyAuth(keys={"sk-my-secret-key"})
agent.serve(host="0.0.0.0", port=8000, dev=True)

agent.serve() starts an HTTP server with registered webhooks, cron triggers, and interfaces in a single process. Add dev=True for hot-reload during development.

Connect to Platforms

from definable.agent import Agent
from definable.agent.interface.telegram import TelegramInterface

telegram = TelegramInterface(bot_token="BOT_TOKEN")

agent = Agent(model="gpt-4o-mini", instructions="You are a Telegram bot.")
agent.serve(telegram)

One agent, multiple platforms. Discord, Desktop, and CLI interfaces also available.

Thinking (Reasoning Layer)

from definable.agent import Agent
from definable.agent.reasoning import Thinking

agent = Agent(
  model="gpt-4o-mini",
  thinking=Thinking(),  # or thinking=True for defaults
  instructions="Think step by step.",
)

output = await agent.arun("What is 127 * 43?")

The thinking layer adds chain-of-thought reasoning before the final response.

Tracing

from definable.agent import Agent
from definable.agent.tracing import Tracing, JSONLExporter

agent = Agent(
  model="gpt-4o-mini",
  tracing=Tracing(exporters=[JSONLExporter("./traces")]),
  instructions="You are a helpful assistant.",
)

output = agent.run("Hello!")
# Traces saved to ./traces/{session_id}.jsonl

Or use tracing=True for default console tracing.

Replay & Compare

from definable.agent import Agent
from definable.model.openai import OpenAIChat

agent = Agent(model="gpt-4o-mini", instructions="You are a helpful assistant.")

# Inspect a past run
output = agent.run("Explain quantum computing.")
replay = agent.replay(run_output=output)
print(replay.steps)  # Each model call and tool invocation
print(replay.tokens)  # Token usage breakdown

# Re-run with a different model and compare
new_output = agent.replay(run_output=output, model=OpenAIChat(id="gpt-4o"))
comparison = agent.compare(output, new_output)
print(comparison.cost_diff)  # Cost difference between runs
print(comparison.token_diff)  # Token usage difference

Replay lets you inspect past runs, re-execute them with different models or instructions, and compare results side by side.

Testing

from definable.agent import Agent
from definable.agent.testing import MockModel

agent = Agent(
  model=MockModel(responses=["The capital of France is Paris."]),
  instructions="You are a geography expert.",
)

output = agent.run("What is the capital of France?")
assert "Paris" in output.content

MockModel returns canned responses — no API keys needed. Use it in unit tests to verify agent behavior deterministically.

---

Features

Category	Details
Models	OpenAI, DeepSeek, Moonshot, xAI, Anthropic, Mistral, Google Gemini, Perplexity, Ollama, OpenRouter, any OpenAI-compatible provider. String shorthand: Agent(model="gpt-4o") resolves automatically
Agents	Multi-turn conversations, structured output, configurable retries, max iterations
Agentic Loop	8-phase pipeline, parallel tool calls via asyncio.gather, HITL pause/resume, cooperative cancellation, EventBus, hooks, ToolRetry, phase metrics
Tools	@tool decorator with automatic parameter extraction from type hints and docstrings
Toolkits	Composable tool groups, KnowledgeToolkit for explicit RAG search
Skills	Domain expertise + tools in one package; 9 built-in skills (incl. MacOS), custom Skill subclass
Knowledge / RAG	Embedders, vector DBs, rerankers (Cohere), chunkers, automatic retrieval
Memory	Session-history memory with auto-summarization
Memory Stores	SQLite, file-based, in-memory
Readers	PDF, DOCX, PPTX, XLSX, ODS, RTF, HTML, images, audio
Reader Providers	Mistral OCR, OpenAI, Anthropic, Google (AI-powered document parsing)
Guardrails	Input/output/tool checkpoints, PII redaction, token limits, topic blocking, regex filters
Guardrails Composition	ALL, ANY, NOT, when() combinators for complex policy rules
Interfaces	Telegram, Discord, Desktop, CLI, session management, identity resolution
Browser Toolkit	55 browser automation tools via Playwright CDP — role-based refs, AI-friendly errors, console/network diagnostics
Claude Code Agent	Zero-dep subprocess wrapper for Claude Code CLI with full Definable ecosystem integration
Runtime	agent.serve(), webhooks, cron triggers, event triggers, dev=True hot-reload
Auth	APIKeyAuth, JWTAuth, AllowlistAuth, CompositeAuth, pluggable AuthProvider protocol
Streaming	Real-time response and tool call streaming
Replay	Inspect past runs, re-execute with overrides, agent.compare() for side-by-side diffs
Middleware	Request/response transforms via agent.use(), logging, retry, metrics
Tracing	JSONL trace export for debugging and analysis
Observability	Live dashboard at /obs/ with real-time events (SSE), session browser, run comparison, tool/model metrics — Agent(observability=True)
Thinking	Chain-of-thought reasoning layer with configurable triggers
Compression	Automatic context window management for long conversations
Testing	MockModel, AgentTestCase, create_test_agent utilities
MCP	Model Context Protocol client for external tool servers
Pipeline	8-phase execution (Prepare → Recall → Think → GuardInput → Compose → InvokeLoop → GuardOutput → Store), custom phases, hooks (before/after/instead)
Debug Mode	Agent(debug=True), DebugExporter with color-coded model call breakdown, DebugConfig for step-mode inspection
Sub-Agents	SubAgentPolicy, spawn_agent tool, concurrent child agents with semaphore-limited execution
Cancellation	CancellationToken for cooperative cancellation, AgentCancelled exception
Types	Full Pydantic models, py.typed marker, mypy verified

Supported Models

from definable.agent import Agent
from definable.model.openai import OpenAIChat        # GPT-4o, GPT-4o-mini, o1, o3, ...
from definable.model.deepseek import DeepSeekChat    # deepseek-chat, deepseek-reasoner
from definable.model.moonshot import MoonshotChat    # moonshot-v1-8k, moonshot-v1-128k
from definable.model.xai import xAI                  # grok-3, grok-2-latest
from definable.model.anthropic import Claude         # claude-sonnet-4-20250514, claude-haiku, ...
from definable.model.mistral import MistralChat      # mistral-large-latest, mistral-small, ...
from definable.model.google import Gemini            # gemini-2.0-flash-001, gemini-1.5-pro, ...
from definable.model.ollama import Ollama            # llama3, mistral, codellama, ...
from definable.model.openrouter import OpenRouter    # any model via OpenRouter
from definable.model.perplexity import Perplexity    # pplx-70b-online, pplx-7b-chat, ...

# Or use string shorthand — no model import needed:
agent = Agent(model="gpt-4o-mini")
agent = Agent(model="anthropic/claude-sonnet-4-20250514")
agent = Agent(model="google/gemini-2.0-flash-001")

Any OpenAI-compatible API works with OpenAIChat(base_url=..., api_key=...). Anthropic, Mistral, Google, and Ollama use their native SDKs (optional deps). Perplexity and OpenRouter use the OpenAI-compatible interface.

Optional Extras

Install only what you need:

pip install definable[readers]          # PDF, DOCX, PPTX, XLSX, ODS, RTF parsers
pip install definable[serve]            # FastAPI + Uvicorn for agent.serve()
pip install definable[cron]             # Cron trigger support
pip install definable[jwt]              # JWT authentication
pip install definable[runtime]          # serve + cron combined
pip install definable[discord]          # Discord interface
pip install definable[browser]          # Browser automation (Playwright CDP)
pip install definable[desktop]          # macOS Desktop Bridge
pip install definable[postgres-memory]  # PostgreSQL memory store
pip install definable[research]         # Deep research (DuckDuckGo + curl-cffi)
pip install definable[mistral-ocr]      # Mistral AI document parsing
pip install definable[mem0-memory]      # Mem0 hosted memory store

Vector DB backends:

pip install definable[pgvector]         # PostgreSQL + pgvector
pip install definable[qdrant]           # Qdrant
pip install definable[chroma]           # ChromaDB
pip install definable[mongodb]          # MongoDB
pip install definable[redis]            # Redis
pip install definable[pinecone]         # Pinecone

Documentation

Full documentation: docs.definable.ai

Project Structure

definable/definable/
├── agent/              # Agent orchestration, config, middleware, loop
│   ├── auth/           # APIKeyAuth, JWTAuth, AllowlistAuth, CompositeAuth
│   ├── compression/    # Context window compression
│   ├── guardrail/      # Input/output/tool policy, PII, token limits, composable rules
│   ├── interface/      # Telegram, Discord, Desktop, CLI integrations
│   ├── observability/  # Live dashboard, metrics, trace browser, SSE events
│   ├── pipeline/       # 8-phase execution pipeline, hooks, ToolRetry, DebugConfig
│   ├── reasoning/      # Thinking layer (chain-of-thought)
│   ├── replay/         # Run inspection, re-execution, comparison
│   ├── research/       # Deep research: multi-wave web search, CKU, gap analysis
│   ├── run/            # RunOutput, RunEvent types
│   ├── runtime/        # AgentRuntime, AgentServer, dev mode
│   ├── tracing/        # JSONL trace export, DebugExporter
│   └── trigger/        # Webhook, Cron, EventTrigger
├── browser/            # BrowserToolkit — 55 tools via Playwright CDP
├── knowledge/          # RAG: embedders, vector DBs, rerankers, chunkers
├── mcp/                # Model Context Protocol client
├── media.py            # Image, Audio, Video, File types
├── memory/             # Session-history memory + 3 store backends (SQLite, File, InMemory)
├── model/              # OpenAI, DeepSeek, Moonshot, xAI, Anthropic, Mistral, Gemini, Perplexity, Ollama, OpenRouter
├── reader/             # File parsers + AI reader providers
├── skill/              # Built-in + custom skills, skill registry
├── tool/               # @tool decorator, Function wrappers
├── toolkit/            # Toolkit base class
├── vectordb/           # Vector database interfaces (7 backends)
└── utils/              # Logging, supervisor, shared utilities

Contributing

Contributions welcome! To get started:

1. Fork the repo and clone it locally
2. Install for development: pip install -e .
3. Make your changes — follow existing code patterns (2-space indentation, 150 char lines)
4. Add tests in definable/tests/ for new features
5. Run ruff check and ruff format for linting
6. Run mypy for type checking
7. Open a pull request

See definable/examples/ for usage patterns.

License

Apache 2.0 — see LICENSE for details.