bottensor-fleet 0.1.0

Graph-native multi-agent fleet for Python. BYO-key. Local-first.

bottensor-fleet

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

---

Demo

---

Install

pip install bottensor-fleet

---

30-second quickstart

from fleet import Graph, Agent

researcher = Agent(
    name="researcher",
    goal="Answer the user's question with citations.",
    model="anthropic/claude-sonnet-4-6",
    tools=["web_search", "web_fetch"],
)

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

import asyncio
from fleet.core.state import GraphState

state = GraphState(goal="What's new in AI safety research this week?")
final = asyncio.run(g.run(state))
print(final.messages[-1].content)

ANTHROPIC_API_KEY=sk-ant-… python examples/solo_agent.py

---

UI

fleet ui

Opens http://localhost:5173. Live graph visualization, agent log, and run controls — all wired to the local WebSocket server.

---

Architecture

┌─────────────────────────────────────────────────────────┐
│  Graph (fluent builder)                                 │
│    .add_node()  .add_edge(cond=...)  .compile()         │
└──────────────────────┬──────────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────────┐
│  Scheduler                                              │
│    BFS walk · asyncio.gather fan-out · state merge      │
│    EventBus pub/sub · per-step checkpoint               │
└──────┬───────────────────────────────┬──────────────────┘
       │                               │
       ▼                               ▼
┌────────────┐               ┌─────────────────────┐
│ Checkpoint │               │  FastAPI + WS       │
│ SQLite /   │               │  /api/runs CRUD     │
│ Redis      │               │  /ws/runs/{id}      │
└────────────┘               └──────────┬──────────┘
                                        │
                                        ▼
                             ┌─────────────────────┐
                             │  React UI           │
                             │  Zustand · ReactFlow│
                             │  Tailwind v4        │
                             └─────────────────────┘

Each node in the graph is a plain async def (GraphState) -> GraphState. Bring your own agents, or use the built-in Agent class for a full ReAct loop.

---

Provider support

Provider	Model prefix	Key env var
Anthropic	anthropic/	ANTHROPIC_API_KEY
OpenAI	openai/	OPENAI_API_KEY
Google Gemini	google/	GEMINI_API_KEY
Groq	groq/	GROQ_API_KEY
Mistral	mistral/	MISTRAL_API_KEY
Ollama (local)	ollama/	(none needed)

Backends are powered by polyrt. Mix providers freely — each Agent picks its own model.

---

Multi-agent patterns

Parallel fan-out

g = (
    Graph("team")
    .add_node("planner",  planner.step)
    .add_node("worker_a", worker_a.step)
    .add_node("worker_b", worker_b.step)
    .add_node("writer",   writer.step)
    .add_edge("planner",  "worker_a")
    .add_edge("planner",  "worker_b")   # fan-out
    .add_edge("worker_a", "writer")
    .add_edge("worker_b", "writer")     # merge
    .set_entry("planner")
    .set_exit("writer")
    .compile()
)

worker_a and worker_b execute concurrently via asyncio.gather.

Conditional cycles

def _needs_fix(s): return not s.scratchpad.get("approved")

g = (
    Graph("review")
    .add_node("reviewer", review_step)
    .add_node("fixer",    fix_step)
    .add_edge("reviewer", "fixer",    cond=_needs_fix)
    .add_edge("fixer",    "reviewer")  # loop
    .set_entry("reviewer")
    .set_exit("reviewer")
    .compile()
)

---

How it compares

Feature	bottensor-fleet	LangGraph	CrewAI	AutoGen
Graph-native	✅	✅	❌ role-based	❌ conversation-first
No framework lock-in	✅ plain Python	⚠️ LangChain	⚠️ CrewAI patterns	⚠️ AutoGen patterns
Multi-provider	✅ polyrt	✅	✅	✅
Built-in UI	✅ React dashboard	❌	❌	❌
Checkpointing	✅ SQLite / Redis	✅	❌	❌
Parallel fan-out	✅ asyncio.gather	✅	❌	⚠️ partial
Conditional cycles	✅ edge cond=	✅	❌	❌

---

Built-in tools

Tool	Description
web_search	DuckDuckGo text search
web_fetch	Fetch a URL, return plain text
python_exec	Execute Python in a subprocess
read_file	Read from FLEET_WORKSPACE
write_file	Write to FLEET_WORKSPACE
list_files	List files under FLEET_WORKSPACE

Custom tools: decorate any async def with @tool.

---

CLI

fleet new <name>          scaffold a new graph project
fleet run <module> <goal> run a graph from the command line
fleet ui                  start the live dashboard
fleet add-agent <name>    scaffold an agent module
fleet ls                  list all recorded runs
fleet replay <run-id>     replay a finished run's events

---

⚠️ Security note

python_exec runs code in an unsandboxed subprocess in v0.1. Only use it with trusted agents and goals. A Docker sandbox is on the v0.2 roadmap.

---

Roadmap

Version	Focus
v0.1 (now)	Core runtime, ReAct agents, FastAPI server, React UI
v0.2	Docker sandbox for python_exec, streaming token output
v0.3	Distributed scheduler (Redis task queue), multi-process workers
v0.4	Persistent vector memory (ChromaDB / pgvector), skill marketplace

---

License

MIT © 2026 Bottensor. See LICENSE.

Built with polyrt · ReactFlow · Zustand · Tailwind CSS · FastAPI