bottensor-fleet

Graph-native multi-agent fleet for Python. BYO-key. Local-first. Ships with a UI.

Why

Most multi-agent frameworks are heavy and locked to one ecosystem. LangGraph is tied to LangChain. CrewAI is opinionated about roles. AutoGen is conversation-first. bottensor-fleet is a small, graph-native runtime that runs anywhere Python runs, lets you bring your own provider key, and ships with a real UI in the wheel.

Install

pip install 'bottensor-fleet[search]'
export ANTHROPIC_API_KEY=sk-ant-...

Extras: [search] adds web tools, [redis] adds Redis checkpointing, [all] gets everything.

30-second example

import asyncio
from fleet import Agent, Graph
from fleet.core.state import GraphState
from fleet.providers.client import FleetLLM

llm = FleetLLM("claude", "claude-sonnet-4-6")
researcher = Agent(name="researcher", llm=llm, tools=["web_search", "web_fetch"])

graph = (
    Graph("solo")
    .add_node("researcher", researcher.step)
    .set_entry("researcher")
    .set_exit("researcher")
    .compile()
)

state = asyncio.run(graph.run(GraphState(goal="What is ReasoningBank?")))
print(state.messages[-1].content)

UI

fleet ui

Opens a local dashboard at http://localhost:8765 with a live DAG view, per-agent logs, token spend, and run history.

CLI

Command	What it does
fleet new <name>	Scaffold a new graph
fleet run <graph.py>	Run a graph from a file
fleet ui	Launch the local dashboard
fleet add-agent	Append an agent to an existing graph
fleet ls	List past runs
fleet --version	Print version

Design

• Graph-native: DAGs with conditional edges and bounded cycles, executed async with asyncio.gather for parallel fan-out.
• BYO-key: Provider abstraction via polyrt. Anthropic and OpenAI in the default install; MLX, Ollama, and others via polyrt extras.
• Checkpointed: Every run persists to SQLite (default) or Redis (opt-in via [redis] extra).
• Tools and skills: @tool decorator auto-derives JSON schemas from type hints. @skill for higher-level capabilities. Web search and fetch built in via the [search] extra.
• UI in the wheel: No separate Node install for users. The React + Vite frontend is bundled into the published wheel.

Comparison

	bottensor-fleet	LangGraph	CrewAI	AutoGen
Graph topology	✅ DAG + cycles	✅	❌ role-based	❌ conversation
Provider-agnostic	✅ via polyrt	⚠️ via LangChain	⚠️	⚠️
Ships with UI	✅	❌	❌	⚠️ Studio (separate)
Pip-install size	~150 KB wheel	heavy	medium	heavy
LangChain dependency	❌	✅ required	❌	❌

Roadmap

• v0.2 — ReasoningBank (Ouyang et al., ICLR 2026): self-evolving agents that learn from successful and failed trajectories. Memory-aware test-time scaling (MaTTS).
• v0.3 — MLX embedder, sequential MaTTS, distributed scheduler.
• v0.4 — Vector memory backend, cloud deploy templates.

Security

The python_exec tool is unsandboxed in v0.1.x. Do not run untrusted graphs. A Docker sandbox lands in v0.2.

License

Apache-2.0. © 2026 Rama Krishna Bachu.

Acknowledgements

Built on polyrt. ReasoningBank design (v0.2) follows Ouyang et al., ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory, ICLR 2026.