The framework for durable agents

$ npx cove@latest init my-agent
Read docs

Like a web framework, but for agents. Markdown for instructions and skills, TypeScript for tools. Durable by default.

An agent is a directory

Define instructions and skills in Markdown, tools in TypeScript, and deploy. The framework compiles the directory, wires up durable workflows, and connects channels.

1 Start with instructions instructions.md

An instructions.md file is a complete agent. Describe its role in Markdown, then run cove.

agent/
instructions.md
# Identity You are a coastal forecast analyst. You can retrieve marine conditions for any harbour in the world.
2 Choose your model agent.ts

cove ships with a default model. Add agent.ts when you want to pick a different one or configure the runtime.

agent/
agent.ts
// pick a model, set the budget export default defineAgent({ model: "harbour-large", maxSteps: 24, })
3 Add reusable skills skills/

Skills are Markdown playbooks loaded only when they are relevant. The agent gets focused guidance without carrying it in every prompt.

agent/skills/
refunds.md
## When to use A customer asks about a charge placed in the last 90 days.

Leverages all Kestrel AI primitives

Model Gateway for inference, Sandboxes, Workflows, and Connect. All critical agent infrastructure works out of the box, no gluing point solutions together.

Runtime

Durable execution, state persistence, event streaming.

Kestrel Workflows

Checkpointed steps, park between messages, resume on delivery.

Model Gateway

Model calls, streaming.

Kestrel Sandbox

Isolated execution.

Kestrel Connect

MCP/HTTP endpoints.

Tools and subagents

Functions, child agents.

Channel

Where your agent gets surfaced.

Chat SDK
RelayBeacon HuddleOrbit Teams Web ChatPigeon APISwitchboard CronTempo

Everything you need for production agents

Enterprise governance, observability, and sandboxed compute come standard. Focus on building, not infrastructure.

Durable execution

Workflows survive crashes and restarts. Every step is checkpointed. Agents park when waiting, resume on the next message.

Sandboxed compute

Agents spin up isolated VMs on demand. File system access, shell execution, and code runs, all completely isolated.

Multi-channel delivery

One agent codebase deploys to web chat, Relay, API, cron, CLI, and your own custom apps.

Human-in-the-loop

Tools that need confirmation trigger approval gates. Sessions park until resolved, then pick up exactly where they stopped.

Subagents

Delegate specialized work to child agents with their own prompts, tools, and sandbox.

Evaluations

Define test suites with scoring rubrics. Run evals on every deployment and on a schedule.

Build your first agent today.

Get started