Inspiration
Investment research is often scattered across chat threads, spreadsheets, links, and disconnected decisions. We wanted the native Codex experience to feel less like a single assistant and more like a disciplined investment team—without giving up the user's control or creating an opaque autonomous-trading system.
What it does
TradingCodex is a thin, local-first investment operating layer on top of native Codex.
- Codex-native investment team: Codex remains in charge of reasoning, research strategy, tools, skills, plugins, and delegation. TradingCodex turns that capability into an adaptive investment workflow rather than a fixed chatbot.
- Adaptive Head Manager: a Head Manager chooses the smallest useful set of investment roles for the question, then returns evidence-backed artifacts, clear handoffs, or an explicit evidence gap.
- Knowledge Wiki: reusable company, product, industry, and technology context remains separate from live evidence and investment conclusions.
- Investment Brain: a versioned, high-freedom layer for hypotheses, causal frames, falsifiers, scenarios, and abstention heuristics.
- Decision Memory: point-in-time decisions, forecasts, outcomes, and lessons can be replayed without allowing history to silently override current evidence.
- Durable boundaries: Django application services own provenance, policy, approval, execution, secrets, and audit records. Research is never treated as an authorization to trade.
How we built it
TradingCodex combines Python 3.11+, Django 5.2, a React/TypeScript viewer, MCP integration, file-native research artifacts, and generated local workspaces. It supports direct optional data access such as OpenBB, while keeping source context, policy checks, and sensitive final effects explicit.
We iterated the product inside Codex with GPT-5.6: shaping the architecture, evolving code and documentation, validating behavior, and producing the demo narrative. Codex accelerated both the engineering loop and the product-design work, while TradingCodex deliberately leaves agent reasoning and tool selection in Codex itself.
Challenges we ran into
The core design challenge was avoiding a second agent platform. Instead of rebuilding a router, permission system, scheduler, or provider framework, we kept native Codex as the intelligence layer and added only the durable boundaries required for investment research and execution safety.
Accomplishments that we're proud of
We created a system that makes Codex feel like a real investment team while retaining clear ownership: flexible intelligence in Codex, reusable context in workspace files, and sensitive actions behind deterministic policy and audit boundaries.
What we learned
In high-stakes research, more autonomy is not automatically better. Durable provenance, explicit evidence gaps, reproducible decisions, and human approval are product features—not afterthoughts.
What's next for TradingCodex
We are expanding reusable Investment Brains and Knowledge Wikis, strengthening workspace workflows, and making evidence-backed investment research easier to run, inspect, and improve over time.
Log in or sign up for Devpost to join the conversation.