Claude Code's headless cousin, now programmable

What it does Flue is a runtime-agnostic TypeScript framework for building autonomous AI agents. It wraps models, sandboxes, tools, and session state into a programmable “harness” — deployable to Node.js, Cloudflare, GitHub Actions, or CI runners. The pitch: it feels like Claude Code, but runs headlessly without a human in the loop.

The interesting bit Most of the agent “logic” lives in Markdown — skills, context, and an AGENTS.md file — while orchestration stays in TypeScript. Sandboxes are tiered: a lightweight virtual one by default (via just-bash), a local() mode for CI that inherits the host shell, or full Linux containers via Daytona when you need real git/Node/browser environments.

Key highlights

• Runtime-agnostic: same agent code runs on Node, Cloudflare Workers, GitHub Actions, GitLab CI/CD, etc.
• Typed outputs: schema-validated results via valibot (or similar) on session.prompt() and session.skill() calls.
• Durable sessions: message history and state persist automatically on supported platforms like Cloudflare.
• MCP adapter: connects to remote MCP servers over streamable HTTP or legacy SSE, keeping secrets in env.
• WebSocket + HTTP surfaces: agents expose /agents/:name/:id endpoints; long-lived instances identified by URL segment.

Caveats

• Explicitly experimental — APIs may change.
• MCP support is limited: no auto-detect, no local stdio spawning, no OAuth callbacks in this version.
• The README’s model strings (anthropic/claude-sonnet-4-6, openai/gpt-5.5) appear speculative or placeholder-like; actual supported models aren’t enumerated.

Verdict Worth a look if you’re already shipping TypeScript and want to turn chat-agent patterns into persistent backend services. Skip it if you need stable APIs today or heavy local MCP tooling — this is early-stage framework code, not a polished SDK.