koda

Koda is the open-source harness for multi-agent, multi-provider AI systems. A control plane you actually use, durable state by default, and every runtime action inspectable from day one.

Install Kodanpx@openkodaai/koda@latestinstall
Read the docs

The platform, day one.

Control plane, runtime evidence, governed memory, skills, evals, channel identity, and release gates — configured from the dashboard, not six config files.

Bot gateway

Channel identity before runtime.

Telegram is the live channel path. Slack and Discord use the same gateway contract in local adapter tests before credentialed rollout.

Telegram
Slack
Discord
TelegramSlackDiscordgateway-gated
Runtime

Multi-agent, multi-provider.

Agents can run alone or inside squad rooms with visible handoffs.

License
Apache
2.0
Fork it, host it, ship it.
Memory & retrieval

Grounded in what's true.

Episodic memory, semantic retrieval, curated relations.

memory
retrieval
curation
relations
Containers

One harness, every container.

Runtime, memory, security — one declarative deploy.

koda-runtime
koda-memory
koda-security
koda-postgres
Multi-model

Multi-provider, multi-model.

Koda verifies native, local, voice, and OpenAI-compatible providers — agents pick what each task needs.

Learn more
/triage-prRun
/summarize-incidentRun
/audit-runsRun
gpt-5
claude-opus
gemini-pro
Atlas
anthropic/claude-opus-4-7
Otto
openai/gpt-5
Scout
google/gemini-3-pro
Pebble
anthropic/claude-haiku-4-5
Sage
anthropic/claude-opus-4-7
Whiskers
ollama/llama3.3
Marlowe
openai/gpt-5
Beacon
google/gemini-3-pro
Echo
anthropic/claude-opus-4-7
Sprout
anthropic/claude-haiku-4-5
Atlas
anthropic/claude-opus-4-7
Otto
openai/gpt-5
Scout
google/gemini-3-pro
Pebble
anthropic/claude-haiku-4-5
Sage
anthropic/claude-opus-4-7
Whiskers
ollama/llama3.3
Marlowe
openai/gpt-5
Beacon
google/gemini-3-pro
Echo
anthropic/claude-opus-4-7
Sprout
anthropic/claude-haiku-4-5
Open foundations

Production‑friendly.
Operator‑first.

Koda treats agent configuration the way operators treat production systems — through a control plane, with inspection, audit trails, and durable state. Programmable where you want it. Steerable where you need it.

Apache 2.0Self-hostableMulti-providerPostgres + S3

No black boxes. No hidden state.

Running a multi-agent system surfaces a hundred operational questions. Koda answers them in its design — not as afterthoughts wrapped around a hardcoded prompt loop.

How do I prevent agent runaway?Which model fits this task?Can I switch providers without rewrites?Where is each agent's state stored?How do I prevent agent runaway?Which model fits this task?Can I switch providers without rewrites?Where is each agent's state stored?How do I prevent agent runaway?Which model fits this task?Can I switch providers without rewrites?Where is each agent's state stored?How do I prevent agent runaway?Which model fits this task?Can I switch providers without rewrites?Where is each agent's state stored?
How do I roll back a bad action?What did the agent see at step 3?Can I run this entirely on my own infra?How do I share a Skill across agents?How do I roll back a bad action?What did the agent see at step 3?Can I run this entirely on my own infra?How do I share a Skill across agents?How do I roll back a bad action?What did the agent see at step 3?Can I run this entirely on my own infra?How do I share a Skill across agents?How do I roll back a bad action?What did the agent see at step 3?Can I run this entirely on my own infra?How do I share a Skill across agents?
Why did the agent retry?What's the audit trail for this run?Where do agents pull memory from?Can I chain agents and inspect each step?Why did the agent retry?What's the audit trail for this run?Where do agents pull memory from?Can I chain agents and inspect each step?Why did the agent retry?What's the audit trail for this run?Where do agents pull memory from?Can I chain agents and inspect each step?Why did the agent retry?What's the audit trail for this run?Where do agents pull memory from?Can I chain agents and inspect each step?

Programmable, not prescriptive

No prescribed agent persona, no locked task domain. Configure providers, agents, and policies the way your system actually needs them.

Production-grade by design

Postgres for durable state. S3-compatible artifact storage. Control-plane OpenAPI plus runtime inspection routes. Built for operators, not hobby projects.

Inspect every action

Every tool call, retrieval hit, and provider response is logged with full context. The runtime state is never hidden behind the prompt.

Yours to fork, yours to run

Apache 2.0 licensed. Docker-first. Self-host on a single VPS or scale across nodes. No managed service, no vendor path dependency.

Start with the platform.

One command to bring up the stack — Postgres, object storage, runtime, and dashboard. Configure providers, agents, and skills from there.

Install Kodanpx@openkodaai/koda@latestinstall
Or read the documentation