What is Doku?

Doku turns one-off agent attempts into governed work that can be run again and checked later. Agents still do the execution. Doku owns the authority around that execution: what is being attempted, what inputs are trusted, what may be accessed, what proof is required, and when a human must decide.

The problem Doku solves

A successful agent chat can produce a useful result, but it rarely leaves a reusable operating model. The instructions, inputs, permissions, decisions, and proof are scattered across the conversation.

Doku gives that work a durable shape:

• a Playbook defines a reusable, outcome-driven recipe;
• a Playbook Run is one governed execution of a published Playbook version;
• Variables declare information the run may need;
• Capacities declare external systems and actions the work may require;
• Claims state what must be true and what evidence proves it;
• human approval stays explicit where an agent should not decide alone.

Doku and the agent

Doku is not the agent and it is not a chat interface. Your agent operates through Doku's MCP and CodeMode surfaces. Doku keeps the run state, permissions, connector custody, evidence, checkpoints, approvals, and audit history.

That separation is deliberate: the agent can reason and act, while Doku remains the source of runtime authority.

The core loop

1. A person chooses a Playbook and supplies the starting context.
2. Doku creates a Run Intent for an authorized agent.
3. The agent claims the work and follows the runnable task graph.
4. Doku records evidence against the Claims that define success.
5. Human-gated Claims pause for an accountable decision.
6. The Run completes only when its required Claims are satisfied.

Where to go next

• Understand the Playbook model
• Author your first Playbook
• Learn how agents execute Runs

Preview status: This documentation surface is an early experiment. The terminology follows Doku's current domain model, but the information architecture and presentation will evolve.