CubePi is a Pythonic, async-native agent framework designed for high performance, readability, and production-grade persistence. It provides a leaner alternative to graph-based agent runtimes by modeling agent logic as a linear while loop that developers can easily trace and debug.

Why CubePi

	langgraph	CubePi
Abstraction	Graph nodes + edges + channels — you model your agent as a state machine	Plain async functions — run_agent_loop is a while loop you can read in 5 minutes
Streaming	Callback-based, multiple handler types	async for event in stream — one pattern everywhere
Checkpointing	Full snapshot per step — serializes entire message list on every channel change	Append-only — writes only new messages, O(1) DB I/O regardless of conversation length
Dependencies	Pulls in langchain-core, langgraph-sdk, and transitive deps	3 core deps: pydantic, anthropic, openai
Tool execution	Tools are graph nodes with manual wiring	Declare tools as functions, framework handles routing and parallel execution
Multi-provider	Via langchain chat model adapters	Native Provider protocol — Anthropic, OpenAI built in, add your own with one class
Middleware	Graph-level middleware on node entry/exit	Agent-level middleware with 8 typed hooks and declarative composition rules
Observability	LangSmith / Langfuse integration, full trace visualization	Native OpenTelemetry — Tracer, Meter, GenAI semconv, OTLP / JSONL exporters built in

Install

pip install cubepi

# Optional extras
pip install cubepi[sqlite]     # SQLite checkpointer
pip install cubepi[postgres]   # Postgres checkpointer
pip install cubepi[mysql]      # MySQL checkpointer
pip install cubepi[mcp]        # MCP tool loaders
pip install cubepi[tracing]    # OpenTelemetry tracing + metrics
pip install cubepi[tracing-otlp]  # Adds the OTLP/HTTP span exporter
pip install cubepi[trace-cli]  # `cubepi trace` terminal viewer

Or with uv:

uv add cubepi
uv add cubepi[sqlite,postgres,mysql,mcp,tracing]

Quick Start

import asyncio
from cubepi import Agent, tool
from cubepi.providers.anthropic import AnthropicProvider

provider = AnthropicProvider(provider_id="anthropic", api_key="sk-...")

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

agent = Agent(
    model=provider.model("claude-sonnet-4-5-20250929"),
    tools=[get_weather],
    system_prompt="You are a helpful weather assistant.",
)

def on_event(event, signal=None):
    if event.type == "text_delta":
        print(event.delta, end="", flush=True)

agent.subscribe(on_event)
asyncio.run(agent.prompt("What's the weather in Tokyo?"))

For a guided tour of the architecture, browse the DeepWiki for this repo or the Core Concepts guide.

Core Concepts

Providers

Abstract LLM interaction behind a Provider protocol. All providers return MessageStream — an async iterator of StreamEvents.

from cubepi.providers.anthropic import AnthropicProvider
from cubepi.providers.openai import OpenAIProvider
from cubepi.providers import FauxProvider

# Real providers
anthropic = AnthropicProvider(provider_id="anthropic", api_key="...")
openai = OpenAIProvider(provider_id="openai", api_key="...")

# Test provider — no API calls, fully deterministic
faux = FauxProvider(provider_id="faux")
faux.set_responses(["Hello!", "How can I help?"])

Use FallbackBoundModel to chain providers — on a rate limit, outage, or context-length error the next model in the chain is tried automatically:

from cubepi import FallbackBoundModel

model = FallbackBoundModel(
    chain=(
        anthropic.model("claude-opus-4-8"),   # primary
        openai.model("gpt-5"),                # fallback
    )
)
agent = Agent(model=model, system_prompt="...")

Tools

Decorate an async function with @tool: the input schema is derived from the typed parameters, the docstring becomes the description, and the framework handles argument parsing, parallel execution, and error wrapping.

from cubepi import tool

@tool
async def search(query: str) -> str:
    "Search the web."
    return f"Results for: {query}"

Need a shared params model, dynamic construction, or execution_mode? The longhand AgentTool(...) is equivalent and fully supported:

from pydantic import BaseModel
from cubepi import AgentTool
from cubepi.agent.types import AgentToolResult
from cubepi.providers.base import TextContent

class SearchParams(BaseModel):
    query: str

async def execute(tool_call_id, params: SearchParams, *, signal=None, on_update=None):
    return AgentToolResult(content=[TextContent(text=f"Results for: {params.query}")])

search = AgentTool(
    name="search",
    description="Search the web",
    parameters=SearchParams,
    execute=execute,
    execution_mode="parallel",  # or "sequential"
)

Middleware

Composable hooks that modify behavior without touching the core loop:

from cubepi import Middleware, compose_middleware
from cubepi.agent.types import BeforeToolCallResult

class LoggingMiddleware(Middleware):
    async def transform_context(self, messages, *, ctx, signal=None):
        print(f"Context has {len(messages)} messages")
        return messages

class SafetyMiddleware(Middleware):
    async def before_tool_call(self, ctx, *, signal=None):
        if ctx.tool_call.name == "dangerous_tool":
            return BeforeToolCallResult(block=True, reason="Blocked by policy")
        return None

hooks = compose_middleware([LoggingMiddleware(), SafetyMiddleware()])

Composition rules:

Hook	Rule
transform_context	Chained — each receives previous result
convert_to_llm	Last implementation wins
resolve_tool_call	First non-None rewrite wins (short-circuits)
before_tool_call	Any block stops execution
after_tool_call	Later overrides earlier
transform_system_prompt	Chained — each receives previous result
after_model_response	Returns TurnAction; last decision wins, messages concatenate
should_stop_after_turn	Any true stops
on_run_end	Messages concatenate; non-empty result triggers one extra model turn

Deferred Tool Groups

When an agent has access to many MCP servers, their combined tool schemas can consume significant context. Deferred tool groups hide schemas by default and let the model expand them on demand:

from cubepi import Agent
from cubepi.deferred import DeferredToolGroup

# load_github_tools is a zero-arg async callable returning list[AgentTool]
# (e.g. wrap load_mcp_tools_stdio(...).tools — see the website guide).
github_group = DeferredToolGroup(
    group_id="mcp:github",
    display_name="GitHub",
    description="Issues, PRs, repos, code search",
    tool_names=["create_issue", "search_repos", "create_pr", "list_comments"],
    loader=load_github_tools,
)

agent = Agent(
    model=provider.model("claude-sonnet-4-6"),
    tools=[get_weather],                     # always-available tools
    deferred_tool_groups=[github_group],      # hidden until requested
)

The model sees a compact catalog in the system prompt instead of full schemas:

# Deferred tool groups

- `mcp:github` — GitHub: Issues, PRs, repos, code search (4 tools)
  create_issue, search_repos, create_pr, list_comments

When the model needs a group, it calls the built-in load_tools tool:

load_tools(group_id="mcp:github")                        # load all
load_tools(group_id="mcp:github", tool_names=["create_issue"])  # or just one

The loader is called once per group per run; subsequent selective loads filter from the cached result. With the default dispatch strategy, load_tools returns the full schemas in its tool result and loaded tools are invoked via the deferred_tool_call dispatcher — the tools array and system prompt stay byte-stable, so loading never invalidates the prompt cache. The v1 behavior (native injection into the model-visible tools array) is available with deferred_tool_strategy="inject".

For advanced use (custom catalog header, cross-run replay), construct DeferredToolsMiddleware directly:

from cubepi.deferred import DeferredToolsMiddleware

# Replay expansion state from a previous run (strategy is required and
# must match the middleware's strategy)
resumed = await DeferredToolsMiddleware.prepare_resumed_state(
    groups=all_groups,
    expanded=saved_extra["expanded_groups"],
    strategy="dispatch",
)
agent = Agent(
    model=model,
    tools=[*builtins, *resumed.pre_loaded_tools],
    deferred_tool_groups=resumed.remaining_groups,
)

Checkpointer

Persist conversation state with append-only semantics:

from cubepi.checkpointer import (
    MemoryCheckpointer,
    SQLiteCheckpointer,
    PostgresCheckpointer,
    MySQLCheckpointer,
)

# In-memory for dev/test
cp = MemoryCheckpointer()

# SQLite for lightweight persistence
async with SQLiteCheckpointer("agent.db") as cp:
    agent = Agent(model=model, checkpointer=cp, thread_id="conv-1")

# Postgres for production
async with PostgresCheckpointer("postgresql://...") as cp:
    agent = Agent(model=model, checkpointer=cp, thread_id="conv-1")

# MySQL for production
async with MySQLCheckpointer("mysql://...") as cp:
    agent = Agent(model=model, checkpointer=cp, thread_id="conv-1")

Postgres and MySQL never issue DDL at runtime — your app owns the schema via Alembic. See the host-integration runbooks (Postgres · MySQL) and the runnable examples/.

FauxProvider for Testing

Ship your agent tests without API keys:

from cubepi.providers import FauxProvider, faux_text, faux_tool_call, faux_assistant_message

provider = FauxProvider(provider_id="faux")
provider.set_responses([
    faux_assistant_message([
        faux_tool_call("search", {"query": "python"}),
    ]),
    faux_assistant_message("Here are the results..."),
])

agent = Agent(model=provider.model("test"), tools=[search_tool])
agent.subscribe(lambda event, signal=None: None)  # subscribe before prompt to receive events
await agent.prompt("Search for python")
# Streams realistic deltas — content_block_start, text_delta, etc.

Tracing

Attach a Tracer and every agent run produces OpenTelemetry spans aligned with the GenAI Semantic Conventions — ingestible by Jaeger, Tempo, Honeycomb, Datadog, AWS X-Ray, or any OTLP-compatible backend without custom instrumentation:

from cubepi.tracing import Tracer, tracing_context
from cubepi.tracing.exporters import JsonlSpanExporter

async with (
    Tracer(
        service_name="my-bot",
        agent_name="assistant",
        exporters=[JsonlSpanExporter(directory="./cubepi-traces")],
    ) as tracer,
    tracer.attached(agent),
):
    with tracing_context(tags=["beta-arm"], metadata={"user_id": "u-42"}):
        await agent.prompt("Hello.")
# On exit: detach (closes any cancelled-run spans + flush) + tracer shutdown.

Span tree per run:

trace
└── invoke_agent  14425.8ms  [0x1cd97cdb]         ← one per agent.prompt()
    ├── cubepi.turn  1283.1ms  [0x5cfda93e]        ← one per LLM round-trip
    │   ├── chat deepseek-v4-flash  1208.7ms  tok 6845/68  [0x0d130229]
    │   └── execute_tool subagent  9610.2ms  subagent  [0x38bdd10a]
    │       └── invoke_agent  9601.0ms  [0x8094f99b]   ← subagent run, nested
    │           └── cubepi.turn  9598.4ms  [0x57c5cfc7]
    │               ├── chat deepseek-v4-flash  1190.3ms  [0x8205ca6b]
    │               └── execute_tool web_search  6500.2ms  web_search  [0xca4e59fc]
    └── cubepi.turn  491.9ms  ERROR  [0xce25f242]
        └── chat deepseek-v4-flash  427.2ms  ERROR  [0x0bff68ec]
            └── error: Error code: 400 - ... `tool_use` ids were found without
                `tool_result` blocks immediately after: call_01_...

No prompts / model outputs are recorded by default. Opt in with Tracer(record_content=True) plus a redact callback for PII. Pair with Meter(...) for gen_ai.client.operation.duration / TTFC / token-usage histograms. Full guide: https://cubepi.ai/docs/guides/tracing/overview

Inspecting traces from the terminal

With JsonlSpanExporter writing to ./cubepi-traces, inspect runs with the cubepi trace CLI (install the extra: pip install cubepi[trace-cli]). All subcommands take --dir (default ./cubepi-traces):

cubepi trace ls                 # recent runs, newest first; the `input`
                                #   column shows the user message + `status`
cubepi trace view <run_id>      # render a run as a tree; errors print inline
                                #   under the failing span (no flag needed).
                                #   A unique run-id PREFIX is enough.
cubepi trace view <run> --content   # also expand prompts / tool args / results
cubepi trace view <run> -v          # expand ALL span attributes (verbose)
cubepi trace follow <run_id>    # stream spans live as they complete
cubepi trace stats --by model   # token / latency / error aggregates
cubepi trace stats --by tool --since 2026-01-01

Typical debugging flow: ls (find the run by its input), then view <prefix> and read the inline error: line under any ERROR span. Need content only recorded with Tracer(record_content=True).

Token / cache fields. The recorder reconciles to the GenAI semconv, so gen_ai.usage.input_tokens is the inclusive total prompt (input + cache_read + cache_creation) and gen_ai.usage.cache_read.input_tokens is a subset of it. From trace fields, cache hit rate is cache_read / input_tokens (≤ 100%) — do not add cache_read to the denominator.

Coding agents debugging cubepi/consumer apps can install the cubepi-trace skill:

npx skills add cubeplexai/cubepi@cubepi-trace -a claude-code

AI Agents

Two skills are available for coding agents (Claude Code, Cursor, Codex, …) working with this repo:

Skill	Install	Purpose
cubepi	npx skills add cubeplexai/cubepi@cubepi -a claude-code	Build agents — API reference, tools, middleware, checkpointing, MCP, HITL
cubepi-trace	npx skills add cubeplexai/cubepi@cubepi-trace -a claude-code	Debug runs — inspect OTel spans, token counts, tool results, streaming failures

Requirements

• Python >= 3.11
• Core: pydantic, anthropic, openai
• Optional: aiosqlite ([sqlite]), asyncpg + sqlalchemy + msgpack ([postgres]), aiomysql + sqlalchemy + msgpack + cryptography ([mysql]), mcp ([mcp]), opentelemetry-sdk ([tracing]), opentelemetry-exporter-otlp-proto-http ([tracing-otlp]), rich ([trace-cli])

Credits

Architecture inspired by pi-agent-core (TypeScript); CubePi is an independent Python reimplementation with Pydantic v2, asyncio-native primitives, and built-in checkpointing.

License

MIT