lovia 0.4.0

A lightweight and elegant Agent framework

lovia

A lightweight, provider-neutral agent framework for Python.

简体中文

import asyncio
from lovia import Agent, Runner

agent = Agent(
    name="Assistant",
    instructions="You are a helpful assistant.",
    model="openai:gpt-4o-mini",
)
result = asyncio.run(Runner.run(agent, "What is the capital of France?"))
print(result.output)  # Paris

Two hard dependencies (httpx, pydantic). No DSL, no graph, no global state. Every advanced feature — tools, sessions, handoffs, structured output, MCP, streaming — is opt-in.

---

Install

pip install lovia

Optional extras:

pip install "lovia[mcp]"    # Model Context Protocol client
pip install "lovia[tools]"  # web_search with DuckDuckGo backend
pip install "lovia[web]"    # FastAPI + SSE chat server

---

Tools

Any typed Python function becomes a tool with @tool. Sync and async both work.

from lovia import Agent, Runner, tool

@tool
def calculate(expression: str) -> float:
    """Evaluate a simple math expression."""
    return eval(expression, {"__builtins__": {}})

agent = Agent(
    name="Calc",
    instructions="Use calculate() for arithmetic.",
    model="openai:gpt-4o-mini",
    tools=[calculate],
)
result = asyncio.run(Runner.run(agent, "What is 1337 * 42?"))

Use Annotated to add per-parameter descriptions to the JSON schema:

from typing import Annotated

@tool
def search(
    query: Annotated[str, "Keywords to search for."],
    limit: Annotated[int, "Max results, 1-20."] = 5,
) -> list[str]: ...

Simple execution policies stay as decorator kwargs:

@tool(timeout=5, retries=2, needs_approval=True)
async def send_email(to: str, body: str) -> str: ...

For advanced cases, pass composable policies; simple kwargs still work.

from lovia import RunContext

async def redact(next_tool, args, ctx):
    result = await next_tool(args, ctx)
    return str(result).replace(ctx.context.api_key, "[redacted]")

@tool(policies=[redact])
async def call_api(ctx: RunContext, path: str) -> str: ...

---

Structured output

Pass any Pydantic model as output_type and the result is validated automatically. output_repair=True lets the model self-correct if the first parse fails.

from pydantic import BaseModel
from lovia import Agent, Runner

class Review(BaseModel):
    rating: int       # 1-5
    summary: str
    pros: list[str]
    cons: list[str]

agent = Agent(
    name="Reviewer",
    instructions="Extract a structured review from the user text.",
    model="openai:gpt-4o-mini",
    output_type=Review,
    output_repair=True,
)
result = asyncio.run(Runner.run(agent, "The battery lasts all day but the screen is dim."))
print(result.output.rating)   # -> int

Override output_type for a single call without touching the agent:

result = await Runner.run(agent, "Summarize in plain text.", output_type=str)

---

Streaming

async for event in Runner.stream(agent, "Tell me a joke"):
    print(event)

Or directly from the agent instance:

async for event in agent.stream("Tell me a joke"):
    print(event)

---

Dynamic instructions

Inject context-aware content at runtime with @agent.system_prompt. Multiple fragments compose with the base instructions.

agent = Agent(name="Support", instructions="You are a support bot.", model="openai:gpt-4o-mini")

@agent.system_prompt
async def inject_user(ctx) -> str:
    user = await db.get_user(ctx.context.user_id)
    return f"The user's name is {user.name}. Their plan is {user.plan}."

# Append one-off context at call time:
result = await Runner.run(agent, "I need help.", append_instructions="Reply in Spanish.")

Prefer functional configuration when cloning reusable agents:

agent = agent.with_system_prompt(inject_user)

---

Handoffs

An agent can delegate to another agent mid-conversation. The Runner follows the chain automatically.

billing = Agent(name="Billing", instructions="Handle billing questions.", model="openai:gpt-4o-mini")
support = Agent(name="Support", instructions="Answer support questions. Hand off billing questions.", model="openai:gpt-4o-mini", handoffs=[billing])

result = await Runner.run(support, "Can I get a refund?")

---

Sessions

Persist conversation history across calls with a session= argument. The default in-memory store is a good starting point; swap in Redis or SQL as needed.

from lovia.stores import InMemorySessionStore

session_store = InMemorySessionStore()

result1 = await Runner.run(agent, "My name is Alice.", session=session_store.session("u42"))
result2 = await Runner.run(agent, "What is my name?", session=session_store.session("u42"))
# → "Your name is Alice."

---

Approval (human in the loop)

Mark sensitive tools with needs_approval=True to require human sign-off.

from lovia import ApprovalChannel

channel = ApprovalChannel()

@tool(needs_approval=True)
def send_email(to: str, body: str) -> str:
    ...

# In your UI, call channel.approve(request_id) or channel.deny(request_id, reason)
result = await Runner.run(agent, "Send a welcome email to [e-mail]", approval_channel=channel)

---

Sync helpers

Runner.run_sync and agent.run_sync are convenience wrappers around asyncio.run. Use them in scripts or wherever you can't await.

result = Runner.run_sync(agent, "What is 2+2?")
print(result.output)

---

Built-in tools

lovia.builtins ships practical, framework-agnostic tools you can drop straight into any agent. Nothing is imported automatically — grab only what you need.

from lovia.builtins.http import http_fetch
from lovia.builtins.search import duckduckgo_search_tool
from lovia.builtins.todo import TodoList, todo_tools
from lovia.builtins.human import HumanChannel, ask_human
from lovia.builtins.think import think
from lovia.builtins.time import now

todos = TodoList()
channel = HumanChannel()

agent = Agent(
    name="Worker",
    instructions="Plan, reason, act.",
    model="openai:gpt-4o-mini",
    tools=[
        http_fetch, now, think,
        duckduckgo_search_tool(),  # requires lovia[tools]
        *todo_tools(todos),
        ask_human(channel),
    ],
)

Builtin convention: stateless helpers export ready-to-use Tool instances, pluggable backends use factories, stateful single-tool helpers expose .tool(), and stateful multi-tool helpers expose .tools().

Filesystem and shell tools live in lovia.sandbox (see next section) — they need a proper sandbox abstraction, not a thin one-off helper.

Runnable demos live in examples/builtins/.

---

Sandbox

lovia.sandbox is the concise filesystem + process layer. One Protocol (Sandbox), one pool (SandboxProvider), one wiring call (attach_sandbox). Ships with a default in-process backend (LocalSandbox); swap in Docker / Firecracker by implementing the Protocol — agent code stays the same.

from lovia import (
    Agent, Runner,
    LocalSandboxProvider, attach_sandbox, AuditStream,
)
from lovia.stores import InMemorySession

base = Agent(name="coder", instructions="…", model="openai:gpt-4o-mini")

async with LocalSandboxProvider() as provider:
    audit = AuditStream()  # optional pub/sub for a UI
    agent = attach_sandbox(base, provider, audit_stream=audit)

    session = InMemorySession()  # from lovia.stores
    await Runner.run(agent, "Create app.py and run it.", session=session, session_id="s1")
    # Same session_id → same workspace on the next turn.
    await Runner.run(agent, "Now add tests.", session=session, session_id="s1")

What you get for free:

• Path traversal guard — symlink-aware, blocks .., /etc/..., etc.
• Dependency isolation by PATH/HOME — each sandbox redirects HOME and TMPDIR to a private subdir and prepends <root>/.venv/bin to PATH. The framework does not manage that venv. When the LLM needs Python deps it bootstraps one itself:

  python -m venv .venv && .venv/bin/pip install pandas
  

  From the next command onwards python and pip resolve to the venv automatically — no special API, no auto-bootstrap, zero pollution of the host environment.
• Audit policy — default_audit_policy() blocks the obvious foot-guns (rm -rf /, mkfs, curl|sh, fork bombs, …) and warns on bare pip install / npm install -g so the LLM is nudged toward a venv. Three-valued (pass/warn/block): warnings annotate stderr without blocking, giving the model a chance to self-correct.
• Per-session lifecycle — sandboxes are refcounted by session_id; multi-turn runs reuse the same workspace (including any .venv the model created) until the provider shuts down.
• Hidden-file filtering — ls/glob skip dotfiles by default so **/*.py doesn't drown in the LLM's own .venv/. Pass include_hidden=True to look.
• Live audit stream — subscribe via AuditStream.subscribe() for a UI; history is kept for late-joining subscribers.
• Apply-patch tool — tolerant unified-diff editor on top of read+write, the cheapest way to let a model edit files.

LocalSandbox is not a security boundary — HOME/PATH redirection keeps things tidy, not safe. For untrusted code use a container-backed Sandbox implementation.

Runnable demos: examples/22_sandbox.py, examples/23_sandbox_session.py, examples/24_custom_sandbox.py.

---

Web UI

lovia.web ships a small FastAPI app + bundled vanilla-JS chat UI. It's the same wiring you'd build yourself, but pre-assembled so you can ship a demo in three lines:

from lovia import Agent
from lovia.web import serve

agent = Agent(name="assistant", instructions="…", model="openai:gpt-4o-mini")
serve(agent, db_path="lovia.db")   # http://127.0.0.1:8000

What you get out of the box:

• Sidebar of chats — every session lives in SQLite (db_path), so it survives restarts, can be renamed, deleted, switched.
• Auto-generated titles — after the first turn a tiny background call asks the same model for a 3-6 word headline.
• Streaming transcript with tool-call cards and approval prompts.
• Workspace panel + audit feed — pass sandbox_provider= and audit_stream= and the right-hand panel shows the per-session files and every shell command's pass/warn/block verdict in real time.

from lovia import LocalSandboxProvider, AuditStream, attach_sandbox

provider = LocalSandboxProvider()
audit = AuditStream()
agent = attach_sandbox(base_agent, provider, audit_stream=audit)
serve(agent, db_path="lovia.db", sandbox_provider=provider, audit_stream=audit)

See examples/25_web_sandbox.py for the full wiring.

---

Skills

Skills are Markdown-driven instruction packs stored in a directory tree. They let you compose domain knowledge without bloating the system prompt.

skills/
  translation/
    SKILL.md          # name, description, usage instructions
    references/       # reference files the agent can read

from lovia.skills import SkillCatalog

catalog = SkillCatalog.from_dir("./skills")   # lazy by default
agent = Agent(
    name="Expert",
    instructions=catalog.render_catalog(),
    model="openai:gpt-4o-mini",
    tools=catalog.tools(),
)

In lazy mode the catalog renders as a compact index; the model calls load_skill to pull in a full skill body on demand. Switch to mode="eager" to inline all bodies up front.

---

Multiple providers

The model= field accepts any "provider:model" string or a Provider instance.

# OpenAI
agent = Agent(model="openai:gpt-4o-mini", ...)
# Anthropic
agent = Agent(model="anthropic:claude-3-5-haiku-20241022", ...)
# Any OpenAI-compatible endpoint
from lovia import OpenAIChatProvider
provider = OpenAIChatProvider(model="deepseek-chat", base_url="https://api.deepseek.com/v1", api_key="...")
agent = Agent(model=provider, ...)

---

Examples

examples/
  01_hello.py                  Minimal agent
  02_tools.py                  Tool calling
  03_streaming.py              Streaming tokens
  04_structured_output.py      Pydantic output
  05_handoff.py                Agent-to-agent delegation
  06_agent_as_tool.py          Sub-agent as a tool
  07_session.py                Persistent sessions
  08_skills.py                 SkillCatalog
  09_compat_provider.py        Custom OpenAI-compatible provider
  10_hooks.py                  Lifecycle hooks / tracing
  11_approval.py               Human-in-the-loop approval
  12_multimodal.py             Image input
  13_budget_and_cancel.py      Token budget & cancellation
  14_guardrails.py             Input/output guards
  15_resume.py                 Resume interrupted runs
  16_web_serve.py              FastAPI + SSE server
  17_responses_reasoning.py    OpenAI Responses API + reasoning
  18_context_policy.py         Auto-summarize long history
  19_dynamic_instructions.py   Dynamic system prompt
  20_builtins.py               Several builtins together
  21_dx.py                     Annotated schemas, run_sync
  22_sandbox.py                Per-run LocalSandbox + sandbox_tools
  23_sandbox_session.py        Multi-turn with LocalSandboxProvider
  24_custom_sandbox.py         Implement Sandbox to plug Docker / firecracker
  25_web_sandbox.py            Full web stack: persistent + sandbox + UI
  builtins/                    One focused demo per builtin
  workflows/                   Multi-agent workflow patterns

---

Development

git clone https://github.com/cymoo/lovia
pip install -e ".[dev]"
pytest          # run tests
ruff check .    # lint
mypy lovia      # type-check

See AGENTS.md for architecture notes, design philosophy, and commit conventions.

---

MIT License