Introduction

Koda is the open-source harness for orchestrating multi-agent, multi-provider AI systems. It is designed to act as a platform you run, not a library you import: you bring the providers, prompts, policies, and task shapes that fit your use case, and Koda supplies the durable operational runtime around them.

What Koda is for

Koda is aimed at teams who need agents running as production software — where every tool call is auditable, state survives restarts, configuration lives in a control plane instead of hand-edited per-agent env files, and the same system can back research workflows, operations automation, support queues, engineering tasks, or any domain-specific agent class.

If you've ever strung together three prompts, two agents, a vector store, and a cron job and wondered where the control plane should live — Koda is that control plane.

Core principles

Five commitments shape how Koda is built. Every feature and default behaviour should be traceable back to one of them.

• Control-plane-first. Providers, agents, secrets, access, and integrations are configured through Koda itself — not through hand-maintained .env files per agent. Bootstrap infrastructure is separate from product configuration, so reverse proxies, VPN overlays, and platform shells stay thin.
• Durable by default. Postgres is the source of truth for control-plane, runtime, knowledge, memory, and audit records. Object binaries flow through a generic S3-compatible contract. Local disk is treated as scratch.
• Inspectable end-to-end. Every provider response, every tool call, every memory write is recorded and reachable through OpenAPI surfaces. There is no hidden state behind the prompt.
• Harness, not monolith. Koda does not force a single agent persona, task domain, or niche. Operators shape agents around the system, not the other way around.
• Operator-first UX. First-run setup, provider wiring, secrets, and agent editing all live in a web dashboard designed for the humans who keep the system running.

Who this is for

If any of the following sounds familiar, Koda is probably the shape you want:

• Teams moving agent workloads off prototype scripts and into something they can keep on-call for.
• Organisations that need self-hosted AI infrastructure — by policy, compliance, data residency, or preference.
• Operators tired of juggling provider credentials, cron triggers, and memory vectors through a YAML tapestry.
• Developers building an agent product on top of a runtime they can actually read the source of and fork.

What you get out of the box

The default installation brings up a coherent stack:

• A Next.js operator dashboard on port 3000, covering setup, provider configuration, agent management, runtime inspection, and operational surfaces.
• A Python control-plane + runtime service on port 8090, exposing OpenAPI-documented routes for every surface the dashboard uses.
• Postgres for durable state — one database, multiple schemas for runtime, control-plane, knowledge, memory, and audit.
• SeaweedFS as the default bundled S3-compatible object store for artifacts and large binaries.
• A doctor command that checks your bootstrap config, storage, secrets, and reachability before you start operating.

Next steps

Pick the path that matches where you are:

• Install Koda — detailed setup with prerequisites, first-run flow, and verification.
• Quickstart — one command to a running agent in under ten minutes.
• Architecture — the mental model: control plane, runtime, knowledge, memory, artifacts.