DISCOVER THE FUTURE OF AI AGENTS

Flue

Added May 4, 2026
Agent & Tooling
Open Source
TypeScriptNode.jsWorkflow AutomationModel Context ProtocolAI AgentsAgent FrameworkAgent & ToolingDeveloper Tools & CodingAutomation, Workflow & RPASecurity & Privacy

A headless TypeScript-based agent runtime framework with built-in multi-tier sandboxes, Markdown-driven logic, and MCP tool integration.

Flue is an agent runtime framework developed by the Astro team (withastro org), positioned as "The Agent Harness Framework." It provides Claude Code-like agent orchestration capabilities but entirely without TUI/GUI, exposed as a pure TypeScript SDK that gives developers fine-grained programmatic control over agent lifecycle, tool injection, sandbox environments, and output structures.

Core features include: a default virtual sandbox based on just-bash (no containers required), a multi-tier sandbox system supporting local filesystem and remote containers (e.g., Daytona), Markdown-driven Skills/Context definitions with auto-discovery, Valibot-based Typed Result Schema validation, Session persistence with sub-agent (Tasks) mechanism, three-level Role override (agent/session/call), Remote MCP Tools support (streamable HTTP and SSE), and a Connector design delivered as Markdown installation instructions.

Flue is runtime-agnostic, deployable to Node.js, Cloudflare (with Durable Objects persistence), GitHub Actions, GitLab CI/CD, and more. Agents can be exposed as HTTP endpoints by declaring triggers = { webhook: true }. Typical scenarios span the full spectrum from simple tool-free agents (e.g., translation) to full coding agents (remote containers + MCP + privileged CLIs).

The project is currently in Experimental status with APIs subject to change. It uses pnpm workspace + Turborepo for monorepo management, with core packages @flue/sdk and @flue/cli, requiring Node.js ≥ 22.

Agent Programming Model#

  • Each agent file exports triggers (trigger declarations) and a default async function (handler) receiving FlueContext (with init, payload, env)
  • Initialization chain: init({ model, sandbox, providers })agent.session()session.prompt() / session.skill() / session.task()
  • Model format: provider/model-name, e.g., anthropic/claude-sonnet-4-6, openai/gpt-5.5

Sandbox System#

  • Virtual sandbox (default): based on just-bash, no containers, suitable for high-traffic scenarios
  • Local filesystem sandbox: sandbox: 'local' mounts the host filesystem
  • Remote container sandbox: via Connector adapters (e.g., Daytona), providing full Linux environments
  • Custom sandbox: supports passing custom Bash factory functions

Tools & Integrations#

  • Privileged CLI injection: defineCommand() injects CLI tools with environment variables on demand (e.g., gh, npm)
  • Remote MCP Tools: supports streamable HTTP (default) and SSE transports for connecting to remote MCP servers
  • Connectors: third-party service adapters delivered as Markdown installation instructions, executed by AI coding agents via flue add <connector>

CLI Commands#

CommandDescription
flue devStart local dev server (default port 3583)
flue run <agent-name>Run a specific agent
flue add <connector>Install a connector

Unconfirmed Information#

  • Whether @flue/sdk and @flue/cli are published to npm public registry
  • Current main branch version number not explicitly labeled
  • Full LLM Provider support list
  • Detailed Cloudflare deployment documentation

Related Projects

View All

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.