AutoHedge

Vibe-Trading stops at research and asks before it touches a broker. AutoHedge doesn’t stop. It splits the job across four agents and runs the whole loop — analysis, decision, risk sizing, execution — against a live Solana wallet.

That makes it the most interesting and the most dangerous page in this category. Read the last section before the first.

The pipeline link

graph LR
    A[Director] --> B[Quant]
    B --> C[Risk Manager]
    C --> D[Execution]
    D --> E[Order]

Each stage emits structured JSON, so the reasoning is inspectable rather than a wall of prose — which matters a great deal when you are trying to work out why it bought something. It’s built on the Swarms agent framework, MIT-licensed.

Run it link

  pip install -U autohedge
  

  from autohedge import AutoHedge

trading_system = AutoHedge(
    name="swarms-fund",
    description="Private Hedge Fund",
)

print(trading_system.run(
    task="Analyze the sentiment of the oil market and provide a thesis on "
         "the overall market position and expected trends."
))
  

There’s also an autohedge CLI. Configuration is a .env:

  JUPITER_API_KEY=          # token price and search
OPENAI_API_KEY=
ANTHROPIC_API_KEY=
WORKSPACE_DIR="agent_workspace"
WALLET_PRIVATE_KEY=""     # read this line twice
  

Solana is the supported venue today, through Jupiter. Coinbase and other exchanges are listed as coming.

What you’re actually signing up for link

• You are putting a wallet private key in a .env file and handing spend authority to an LLM pipeline. Nothing between the thesis and the order asks you to confirm. Use a fresh wallet funded with an amount whose total loss would be merely annoying, and never a key that also holds anything else.
• The README’s claims run ahead of its documentation. “Enterprise-grade” and “institutional reliability” describe the intent; what’s documented is four agents, an env file, and one example. There is no documented dry-run or paper mode, no published backtest, and no stated risk limits you can configure — the “risk-first architecture” is an agent asked to think about risk, not a hard cap the code enforces.
• Check the commit dates before you trust it with money. As of this writing the last push was several months back. In a repository that executes trades, staleness is a material fact, not a detail.
• An LLM cannot know whether an edge is real. The pipeline will produce a confident thesis for anything you point it at, including noise. Confidence is the output it is best at.
• Autonomous trading may not be legal for you to run as-is. Managing other people’s money, in most jurisdictions, is a licensed activity. Your own funds are your own business; the moment anyone else’s are involved, that’s a question for a lawyer, not a README.

Where it fits link

The genuinely useful thing here is the architecture: four narrow agents with one job each, structured hand-offs between them, and a risk stage that exists as a separate step. That pattern is worth borrowing whatever domain you work in. Borrowing the pattern costs nothing. Running the loop costs whatever is in the wallet.

Next link

None of that matters if you can’t tell whether the output is any good → promptfoo