Fraktál
Fable 5 × Grok 4.5
A high-performance multi-agent coding engine. Combines elite reasoning architecture with a battle-tested build pipeline — powered by DeepSeek API for professional-grade output at a fraction of frontier-model cost.
Two Paradigms, One Framework
Fraktál merges two elite AI coding architectures into a single, cohesive agent — delivering high-performance reasoning at commodity-model cost.
Fable 5 Reasoning
Hypothesis-driven thinking, evidence-based claims, adversarial self-verification. Multi-provider LLM orchestration, persistent memory, and MCP server for Reasonix interop.
Grok 4.5 Build Pipeline
Battle-tested plan → design → implement → review workflow. Agent, persona, role, and skill architecture proven in Grok Build — now powered by DeepSeek.
Specialised Outputs
Seven domain playbooks guide every build — dashboards, reports, websites, datasets, research, presentations, and infographics. Every output meets H Heuristics professional standards.
High Performance, Low Cost
DeepSeek-first architecture — 10–20× cheaper than frontier models for equivalent quality. Fast inference, 128K context, and chain-of-thought reasoning via DeepSeek R1.
Plan · Design · Implement · Review
The flagship Grok Build workflow — four phases handled by specialised sub-agents, with DeepSeek API powering every step.
Plan
Architect explores the codebase, identifies patterns, and produces a step-by-step implementation plan.
Design
Design-doc writer produces DESIGN.md. Reviewer audits it. Writer revises until zero open issues.
Implement
Coder writes code matching existing patterns. Verifier checks correctness. Loop until clean.
Review
Final review against the diff. Security audit on high effort. The gate before shipping.
Everything Required to Ship
Nine agent roles, eight personas, five pipeline skills, seven domain playbooks, and eight tools — one Python package.
Multi-Agent Orchestration
Architect, Coder, Verifier, and Reporter sub-agents — each with specialised system prompts and tools. The Orchestrator delegates and synthesises.
Persistent Memory
SQLite-backed memory survives across sessions. The Orchestrator recalls past decisions and task outcomes automatically on each run.
MCP Server
Expose all Fraktál tools to Reasonix and any MCP-compatible AI agent via stdio. Drop-in interop with your existing workflow.
Domain Playbooks
Seven playbooks guide specialised outputs — dashboards, reports, websites, datasets, research, presentations, and infographics.
Eight Built-in Tools
Read, write, list files. Execute shell commands. Regex code search. Plus H Heuristics domain tools — dashboard validation, data sources, market sizing.
DeepSeek-First, Multi-Provider
Default: DeepSeek V3 + R1. Also supports OpenAI, Anthropic, Ollama, and LM Studio via the provider registry.
Get Fraktál
Download the latest release directly from fraktal.hheuristics.com. Extract anywhere and install. No git, no GitHub — just the code.
Requires Python 3.10+ and a DeepSeek API key.
Set with: export DEEPSEEK_API_KEY="sk-your-key-here"
Running in Under a Minute
Three steps from zero to your first Fraktál build.
Install & Configure
Download the archive above, extract, and run pip install -e .. Set your DEEPSEEK_API_KEY environment variable. Run fraktal setup to initialise.
Verify Connection
Run fraktal health to confirm API connectivity. Check available models with fraktal models. Browse agents, personas, and skills.
Run Your First Build
Execute fraktal build "your task" for the full pipeline — plan, design, implement, review. Or use individual commands for specific phases.
Fraktál vs. the Alternatives
Fraktál combines the best of both worlds — and costs a fraction of the price.
| Feature | Fable 5 | Grok Build | Fraktál |
|---|---|---|---|
| LLM Backend | Multi-provider | Grok 4.5 (xAI) | DeepSeek-first, multi-provider |
| Build Pipeline | Domain playbooks | plan → design → implement → review | Grok Build + playbooks |
| Agent Model | Orchestrator + sub-agents | Agent / Persona / Role | Both — unified architecture |
| Persistent Memory | ✓ SQLite / JSON | — | ✓ SQLite / JSON |
| MCP Server | ✓ | — | ✓ Reasonix-native |
| Domain Playbooks | ✓ 7 topics | — | ✓ 7 topics |
| Approx. Cost per Task | Provider-dependent | Premium (Grok API) | ~10–20× cheaper (DeepSeek) |
| License | MIT | Proprietary | MIT |