Unified Broker Interface (UBI)
UBI is the single, normalized interface across brokers. Each broker is reached through its own Model Context Protocol (MCP) server, and UBI presents them to agents and researchers as one consistent set of tools, so research code never depends on one broker’s API.
- One adapter per broker MCP server; broker-specific details stay behind the interface.
- Read-oriented by default: market data, reference data, and — where applicable — simulated (paper) account data.
- Credentials stay with each MCP server and are never exposed to models.
- Any order-capable tool is separate, off by default, confined to paper accounts in the research stage, and gated by the policy engine.
Specialized MCP servers
Every data source or tool sits behind its own MCP server. This isolates credentials and permissions, gives each capability a narrow, typed interface, and makes access auditable.
- One server per concern (for example market data, filings and news, and reference data).
- Read-only by default; any write capability is a separate, explicitly permissioned server.
- Typed inputs and outputs, with rate limits and request logging.
Python agent harness
The harness runs role-specialized agents against MCP tools. It owns scheduling, budgets, schema validation, retries, and recording.
- Agents receive only the tools their role needs.
- Interactive sessions through an MCP-capable chat client, alongside scheduled batch runs.
- Outputs are validated against schemas before anything else sees them.
- Every prompt, response, tool call, and model version is recorded for replay.
Deterministic policy engine
Rules are code. The engine evaluates structured proposals and returns approve, reject, or escalate — with the same answer every time for the same input.
- No model in the decision path; the engine is unit-testable and auditable.
- Limits, allow-lists, and kill switches are configuration reviewed like code.
- Every decision is logged with the rule that produced it.
AWS integration
The target deployment uses managed AWS services so the environment stays observable, least-privilege, and reproducible.
- S3 for versioned datasets and run artifacts; encryption at rest throughout.
- Scheduled and batch compute (for example Lambda or containerized jobs) with per-component IAM roles.
- Secrets Manager for credentials; CloudWatch for logs, metrics, and alarms.
- Infrastructure defined as code and changed through review.