withastro/flue

What it solves

Flue provides a programmable TypeScript harness for building autonomous agents that go beyond simple chatbots. It solves the problem of providing agents with the necessary environment—including secure sandboxes, persistent sessions, and integrated tools—to perform complex, multi-step tasks autonomously.

How it works

Flue allows developers to define agents as functions where they can compose a "harness" of resources. By using hooks like useModel, useSandbox, useSkill, and useTool, developers can specify the model, the secure execution environment, and the specific capabilities (skills and tools) the agent needs. These agents can be run locally via a CLI or deployed to various environments like Node.js, Cloudflare Workers, or GitHub Actions.

Who it’s for

TypeScript developers building autonomous AI agents that need to interact with the filesystem, call APIs, and maintain state across conversations and events.

Highlights

• Secure Sandboxes: Provides environments where agents can safely modify files and execute tools.
• Durability: Ensures agents preserve progress through failures and restarts.
• Subagents: Supports the delegation of tasks to specialized expert roles.
• Skills and Tools: Allows packaging reusable expertise (skills) and and typed actions (tools) for agents to use.
• MCP Support: Connects to the Model Context Protocol ecosystem for authenticated tools and services.
• Observability: Integrates with OpenTelemetry, Braintrust, and Sentry for monitoring and telemetry.
• Multi-platform Deployment: Deployable to Node.js, Cloudflare Workers, and CI/CD pipelines.