Skip to content

Kohakku

Agent dispatch platform — orchestrate AI agents across Docker, Kubernetes, and ECS.

Kohakku is the execution layer in the CONFLICT ecosystem. It takes a task specification, assembles a brief with skills and configs, dispatches an agent container to the right infrastructure, and manages the full lifecycle from spawn to completion.

What it does

  • Dispatch agents to Docker, Kubernetes, or AWS ECS with a single API call
  • Assemble briefs from skills, configs, and instructions — agents get everything they need at boot
  • Orchestrate workflows via Temporal — chained, fan-out, supervisor/worker, review loops
  • Multi-tenant — organizations, teams, workspaces, projects with RBAC
  • Observe everything — dashboards, charts, log streaming, progress reporting

Quick Start

git clone git@github.com:ConflictHQ/kohakku.git
cd kohakku
make setup && make up && make migrate && make seed

Controller: http://localhost:8000 — Dispatcher: http://localhost:8080

Architecture

Controller (Django)          Dispatcher (Go)           Agents (Docker)
┌──────────────────┐        ┌──────────────────┐      ┌──────────────┐
│ Task management  │───────▶│ Queue consumer   │─────▶│ agent-basic  │
│ Skill registry   │        │ Container spawn  │      │ agent-code   │
│ Brief assembly   │        │ Status monitor   │      │ agent-term   │
│ Workflow engine  │        │ Agent protocol   │      │ agent-browser│
│ Tenant/RBAC      │        │ Log streaming    │      │ agent-desktop│
└────────┬─────────┘        └──────────────────┘      │ agent-full   │
         │                                             └──────────────┘
    ┌────┴────┐     ┌─────────┐     ┌──────────┐
    │Postgres │     │  Redis  │     │ Temporal │
    │ + MinIO │     │         │     │          │
    └─────────┘     └─────────┘     └──────────┘

Key Concepts

Concept Description
Task A unit of work dispatched to an agent container
Brief Assembled package of skills, configs, and instructions given to an agent at boot
Skill Reusable capability package (zip with metadata, instructions, scripts)
Workflow Multi-step orchestration pattern powered by Temporal
Dispatcher Go service that spawns containers and manages agent lifecycle
Agent Tier Image capability level — basic, codeready, mesh, terminal, browser, desktop, full
Runtime AI runtime inside the agent — claude, codex, gemini, or all

Features

  • Three runtime backends — Local Docker, Kubernetes (Jobs), AWS ECS (Fargate)
  • Seven agent tiers — from CLI-only to full desktop with VNC
  • Six workflow patterns — single dispatch, chained, fan-out, supervisor/worker, review loop, advisor
  • Brief pipeline — skill upload, versioning, dedup, presigned URLs, hash verification
  • Auth stack — django-guardian RBAC, OAuth2, scoped API keys, HMAC-signed callbacks
  • Secrets management — Fernet encrypted local store, AWS Secrets Manager, Vault backends
  • Operator dashboards — task status charts, agent gallery with noVNC, telemetry, workflow progress
  • Infrastructure as code — Terraform modules for ECS, EKS, GKE; Helm charts for Controller and Dispatcher