Loki Mode

Multi-Agent Autonomous Startup System

From PRD to deployed product with zero human intervention. Research-backed patterns from Anthropic, DeepMind, and OpenAI.

View on GitHub Explore Documentation

Download PPTX for offline viewing

Anti-Sycophancy

CONSENSAGENT blind review system with 3 parallel reviewers + devil's advocate to prevent groupthink.

ACL 2025

Quality Gates

7-gate system tracking velocity-quality balance. Zero tolerance for complexity increase.

arXiv 2511.04427v2

Scale Patterns

Recursive sub-planners, judge agents, and optimistic concurrency for 100+ agent scale.

Cursor 2026

Full SDLC

37 specialized agents covering engineering, operations, business, and growth.

Complete

Memory Systems

Episodic, semantic, and procedural memory architecture for continuous learning.

A-Mem Research

Constitutional AI

Self-critique against principles with debate-based verification for critical changes.

Anthropic

Realtime Dashboard

Kanban board, agent cards, RARV cycle, quality gates. Anthropic design language.

v4.0.0

CLI + Distribution

loki CLI wrapper, YAML config, npm/Homebrew/Docker. Install via brew, npm, or docker.

v4.1.0

Foundational Principles

WHY-based reasoning: Autonomy, Memory > Reasoning, Verification, Atomicity, Constraints.

v4.2.0

CoVe + MemEvolve

Chain-of-Verification anti-hallucination. Task-aware memory retrieval. Research-backed quality.

v5.2.4 NEW

Multi-Provider Support

Claude Code, OpenAI Codex CLI, Google Gemini CLI. Abstract tiers: planning, development, fast.

v5.0.0

Human Intervention

PAUSE, HUMAN_INPUT.md, STOP files. Ctrl+C pause/exit. Full control when needed.

v3.4.0

AI-Powered Merge

Claude automatically resolves git merge conflicts. Intelligent code integration.

v3.4.0

Complexity Tiers

Auto-detect simple/standard/complex. 3/6/8 phases based on project scope.

v3.4.0

Quick Start

# Launch Claude Code with autonomous permissions
claude --dangerously-skip-permissions

# Then invoke:
# "Loki Mode" or "Loki Mode with PRD at path/to/prd"

Full README

Complete installation and usage guide

Realtime Dashboard

Monitor and manage Loki Mode with a production-ready web interface

Total Tasks

12

Active Agents

4

Completed

8

Phase

DEVELOPMENT

AUTONOMOUS ACT 3/6 Gates Passed

Kanban Board

4-column task queue: Pending, In Progress, Review, Completed. Drag-and-drop for local tasks.

Agent Cards

Live status for all 37 agent types. Model badges (Opus/Sonnet/Haiku), runtime stats.

RARV Cycle

Visual step indicator: Reason - Act - Reflect - Verify. Updates in realtime.

Quality Gates

6 gates with pass/pending/fail icons. Static analysis, 3-reviewer, anti-sycophancy.

Memory System

Progress bars for episodic (traces), semantic (patterns), procedural (skills) memory.

Theme Support

Light/dark mode with Anthropic design language. System preference detection.

Dashboard Guide

Complete documentation for all UI components, features, and architecture

Launch Dashboard

# Start local server in .loki directory
cd autonomy/.loki
python3 -m http.server 8080

# Open dashboard
open http://localhost:8080/dashboard/index.html

Architecture & Core Workflow

Foundational patterns and system design

Core Workflow

RARV cycle: Reason → Act → Reflect → Verify

Agent Types

37 specialized agents in 7 swarms + orchestration

SDLC Phases

Bootstrap → Discovery → Architecture → Development → Deployment → Growth

Memory System

Episodic, semantic, and procedural memory

Quality Control

7-gate system with anti-sycophancy protocols

Tool Orchestration

Efficiency metrics and reward signals (ToolOrchestra)

Task Queue

Queue management, circuit breakers, dead-letter handling

Cursor Scale Patterns

100+ agent coordination, recursive sub-planners, judge agents

MCP Integration

Model Context Protocol server capabilities

Research Foundation

Built on cutting-edge research from leading AI labs

OpenAI Patterns

Agents SDK, guardrails, tripwires, handoffs, tracing

Lab Research

DeepMind + Anthropic: Constitutional AI, Scalable Oversight via Debate

Production Patterns

Hacker News 2025: What actually works in production

Advanced Patterns

OptiMind, k8s-valkey-operator, Multi-Agent Reflexion

Business Operations

Growth, marketing, sales, finance agents

Deployment

Cloud deployment, CI/CD, monitoring

Confidence Routing

4-tier routing: auto-approve, direct, supervisor, escalate

Prompt Repetition

arXiv 2512.14982v1: 4-5x accuracy improvement for Haiku

Acknowledgements

25+ research papers, Anthropic, DeepMind, OpenAI sources

Skill Modules

Progressive disclosure: Load modules based on your current task

Module Index

Selection rules: Load 1-3 modules based on your task

Model Selection

Opus/Sonnet/Haiku usage, Task tool, thinking modes

Quality Gates

7-gate system, blind review, velocity-quality balance

Testing

Playwright MCP, property-based testing, E2E automation

Production

HN patterns, confidence routing, context management

Parallel Workflows

Git worktrees, optimistic concurrency, inter-stream coordination

Advanced Patterns

OptiMind, formal state machines, Constitutional AI

Agent Orchestration

Structured prompting, handoffs, routing mode

Artifacts

Code transformation, report generation

Troubleshooting

Common issues, red flags, fallback systems

Multi-Provider Support

NEW in v5.0.0: Run Loki Mode with Claude Code, OpenAI Codex CLI, or Google Gemini CLI

Feature	Claude Code	OpenAI Codex	Gemini CLI
Full Features	Yes	Degraded	Degraded
Task Tool (Subagents)	Yes	No	No
Parallel Agents	10+	No	No
MCP Integration	Yes	No	No
Context Window	200K	128K	1M
Model Tiers	opus/sonnet/haiku	effort param	thinking param

Claude Code (Default)

Full-featured provider with Task tool, parallel agents, and MCP support

claude --dangerously-skip-permissions

OpenAI Codex CLI

Degraded mode with effort parameter. Sequential execution only.

--provider codex

Google Gemini CLI

Degraded mode with thinking_level. Largest context window (1M).

--provider gemini

Provider Selection

# Via CLI flag
./autonomy/run.sh --provider codex ./prd.md
loki start --provider gemini ./prd.md

# Via environment variable
export LOKI_PROVIDER=codex
loki start ./prd.md

Full Provider Documentation

Complete provider comparison, configuration, and usage guide

Examples & Demos

Real PRDs and demo projects

Simple Todo App

Basic CRUD with React + Express

API Only

REST API with authentication

Full Stack Demo

Complete application with frontend, backend, and database

Static Landing Page

Marketing site with analytics

Demo Guide

Recording and running demos

Installation

Setup guide and prerequisites

Comparisons & Analysis

Factual technical comparisons with other systems

Loki Mode vs Auto-Claude

Honest technical comparison with top competitor (9,594 stars)

Loki Mode vs Cursor

Scale patterns, quality gates, proven results comparison

Multi-Agent Systems

Comparison with other agent frameworks

Competitive Analysis

Auto-Claude, MemOS, Dexter analysis with 2026 trends

Thick-to-Thin Refactor

v3.0 progressive disclosure architecture

Changelog

Complete version history from v1.0.0 to v4.2.0

Constitution

Core principles, autonomy rules, guardrails