Early access

Any model, any agent, one governed workflow.

Yakusha is a visual harness builder for AI agents. Bring any LLM, including Codex and Claude, give each agent its own identity, and reuse agents across workflows with designed handoffs. Resource-level governance keeps every agent to what it's allowed to touch.

Join the waitlist See how it works

役者 (yakusha) means “actor” in Japanese. Your agents are the cast. You direct.

Built for indie builders

Design agent systems visually, not in glue code

Lay out your agents, how work moves between them, and what you see at the end, all in one place.

Visual harness builder

Build harnesses for your agents on a canvas. Each harness becomes a workflow you can run again and again.

Agent-native workflows

Unlike traditional workflows, Yakusha workflows are built around agents. Create an agent once and reuse it across workflows.

An identity for every agent

Each agent gets its own identity, so you always know who did what across every workflow it plays a part in.

Bring any LLM

Pick the model that drives each agent, including coding agents like Codex and Claude. No lock-in to one provider.

Native interfaces

Put a button on a workflow: click it, get results. Build simple views that display the data your workflows store.

How it works

Cast, direct, perform

A systematic way to direct workflows and the handoffs between agents.

01 · Cast

Agents

Create agents with their own identity and choose the LLM that drives each one.
02 · Direct

Workflows & handoffs

Arrange agents into workflows visually and define exactly how work is handed from one to the next.
03 · Perform

Interfaces

Trigger workflows from a button and view the stored results in a native interface.

For teams that need guardrails

Governance built in, not bolted on

AI overreach happens when an agent reaches past its job. Yakusha enforces strict, resource-level governance, so every agent stays in its role.

Resource-level permissions. Agents only touch the resources they're allowed to.
Distinct agent identities. Access is tied to who the agent is, not to a shared key.
Designed handoffs. Work moves between agents the way you defined it, not however a model improvises.

# illustrative example
agent: researcher
driver: any LLM
resources:
  docs/*      read  ✓
  repo/main   write ✗
handoff: → writer

Early access

Join the waitlist

Be among the first to build with Yakusha. We'll email you when it's ready.

No spam. Just a note when it's ready.