Overview

Introduction

What Clawlas is — a personal AI research laboratory for speculative crypto strategies — and the stack it runs on.

Clawlas is a personal AI research laboratory for discovering, evaluating, validating and executing highly speculative crypto strategies around new listings. It is not a trading bot and not an autonomous AI making uncontrolled financial decisions: a team of specialized agents proposes with measurable evidence, and the human governs every move. This documentation is the working reference for the current stack: a dual TypeScript + Python architecture — a TanStack Start + Convex web layer and a Python research/execution core (LangGraph, FastAPI, the quant stack) — with SQLite memory and Parquet/DuckDB analytics, all containerized on a single VPS.

Status: working draft

The specs are not finalized — this site exists to illustrate the whole surface of what will be built, page by page, so decisions are visible before code lands. Initial scope: new crypto listings (Binance, Coinbase, Bybit, Hyperliquid, OKX, KuCoin, MEXC, DEX launches, Pump.fun, Solana ecosystem), shadow-mode exploration on live data.

The one question it is built to answer

The goal is not to predict the future. It is to continuously answer one question:

Given everything currently known…

“…what is the statistically best decision right now?” Every decision is explainable, measurable, reproducible, challengeable, and continuously improved.

The vocabulary

The terms are deliberate and used consistently across every page:

Name	What it is
Clawlas	The laboratory as a whole — collectors, data backend, models, scoring engine, execution, web UI, infrastructure.
The agents	The multi-agent research system — scout, research, quant, security, on-chain, social, strategy, backtesting, debate, risk, memory, execution — running as LangGraph workflows in the Python core.
The human	The governor — owns the strategy and validates deployment, capital, execution permissions and scoring. The platform proposes; the human governs.

Key design decisions

Clawlas commits to full self-hosting to maximize control and learning of the fundamentals — the major decisions:

Topic	Decision
AI / agent layer	Python multi-agent system (LangGraph) + quant core
Data backend	Self-hosted Convex (Docker, on the VPS)
Agent memory	Dedicated SQLite, local to the runtime container
Deployment	Everything containerized on the VPS (Docker + Caddy + CI/CD)
Evaluation	Shadow mode on live data (no-real-money key), after backtest
Journal	Immutable trade & decision journal (reasoning frozen at time T)
Asset classes	New crypto listings now, designed to add Forex / commodities as profiles

The discipline stays intact

Deterministic execution, hard-coded guardrails, backtest first, everything logged. The agents act only inside the guardrails and only within human-granted permissions — arbitrary orders stay impossible, and during research the no-real-money key is the physical safety. The Python core inherits that discipline — it does not dilute it.

How to read these docs

Each section maps to a slice of the system you can reason about in isolation:

The shape

Architecture

The research pipeline, the Docker topology, the Convex ↔ SQLite data boundary, and the multi-profile seams frozen now to add Forex later without a rewrite.

The brain

The research system

The multi-agent system, the adaptive scoring engine, continuous re-evaluation, the kernel and orchestration, and human governance.

The truth

Memory & journal

What Convex stores, the SQLite memory layer, and the immutable trade & decision journal — the heart of the research.

The discipline

Phases & operations

Shadow mode, the guardrail layer, the research loop, libraries, and the CI/CD pipeline.

The research process

Observe new listings on live market data, in real conditions, risking nothing.
Freeze every decision and its reasoning at the instant it was made.
Compare with reality across horizons and answer honestly: real edge, or just narration?