Features
Train reinforcement learning trading agents and deploy them for AI live trading. One platform for experiments, monitoring, execution, and MCP-connected AI hosts.
Training & experiments
Configure environments and policies, launch reproducible training runs, and monitor progress in real time. Every run is versioned and tracked.
Create experiments with full control over RL environment, policy, actions, observations, and rewards. Validate before you run, then watch key metrics and failures stream live from the dashboard.
Learn more about training · Reinforcement learning trading guide · Terms you’ll seeLive trading deployment
Promote a trained model to a capital deployment on your broker. Paper and live share one path; live is quality-gated until production ready. Cloud schedule, no local setup.
Choose a model, link a broker account, and create a deployment. Timeframe and universe come from your experiment. The platform runs inference and places orders on schedule, aligned to market hours—same pipeline for paper and live.
Learn more about live tradingBroker & account flexibility
Use one broker for both data and execution. Create an account through us or add external broker accounts (Alpaca, Forex.com, and others as available).
Experiments and live deployments both use the same broker account—so training data and live execution stay consistent. Credentials are stored securely; only secret references live in the database.
Learn more about broker accountsCLI
Terminal RL agent ops: train, trade, and automate from your shell. Fullscreen app, subcommands, and optional --prompt—not a coding agent or second dashboard.
Install the kabu binary from GitHub Packages, sign in with browser OAuth, and operate training runs and live deployments from the terminal or CI.
MCP for AI agents
Connect Claude Desktop, Cursor, Grok Build, or any MCP host to Kabu over HTTPS. Interactive tables and charts where the host supports MCP Apps—without putting broker secrets in the model.
Point the host at https://mcp.usekabu.app. OAuth Connect (URL only) — login and pick a workspace in the browser. Compatible hosts can click from an experiment into its broker or runs without leaving the agent UI.