cayu-dev/cayu

What it solves

Cayu is designed for building production-grade, long-horizon agents that must handle consequential work. It addresses the "boundary failures" common in agent prototypes—such as process crashes after side effects, incorrect tool authority, the need for human-in-the-loop approvals, and the loss of cost or causal attribution during complex runs.

How it works

Cayu separates the agent's identity from its execution resources and state. It provides a Python framework to assemble a "harness" (context, tools, memory, and policies) and an integrated durable runtime that manages sessions, transcripts, and checkpoints. This ensures that agent work can survive process restarts and be resumed from the last known state.

Who it’s for

It is intended for developers building domain-specific AI agents that require high reliability, durable state management, and strict control over the execution lifecycle, particularly for tasks involving multi-step workflows and human intervention.

Highlights

• Durable Runtime: Built-in machinery for sessions, checkpoints, and resumable workflows that survive process boundaries.
• Comprehensive Tooling: Support for typed Python tools, Model Context Protocol (MCP), and stateful browser tools.
• Human-in-the-Loop: Explicit support for approval gating and manual recovery checkpoints.
• Multi-Agent Orchestration: Capabilities for subagents, delegation, and durable peer requests.
• Production Readiness: Integrated usage accounting, cost attribution, and a FastAPI control plane with a dashboard for monitoring sessions and evals.