agent-bundle

Define agent skills in YAML. Generate a production-ready agent service.

Anthropic · OpenAI · Gemini · Ollama · OpenRouter — E2B · Kubernetes sandboxes — AWS ECS Fargate deploy (beta)

The Pipeline

1. Define — one YAML file describes your entire agent:

# agent-bundle.yaml
name: personalized-recommend

model:
  provider: openrouter
  model: qwen/qwen3.5-397b-a17b

prompt:
  system: |
    You are a personalization assistant.
    Use the filesystem tools to read/write user profile memory.
    Recommend relevant products with concise reasons.

sandbox:
  provider: e2b
  e2b:
    template: base
    build:
      dockerfile: ./Dockerfile

mcp:
  servers:
    - name: fs
      transport: stdio
      command: npx
      args: ["@modelcontextprotocol/server-filesystem", "/data", "/memory"]

skills:
  - path: ./skills/update-memory
  - path: ./skills/recommend

2. Generate — Prisma-style typed codegen:

npx agent-bundle generate

node_modules/@agent-bundle/personalized-recommend/
├── index.ts      ← typed agent factory
├── types.ts      ← type definitions
├── bundle.json   ← resolved config snapshot
└── package.json  ← scoped package metadata

3. Use — import the generated package and embed it in your own service:

// server.ts — your own Hono / Express / Fastify app
import { PersonalizedRecommend } from "@agent-bundle/personalized-recommend";

const agent = await PersonalizedRecommend.init();

app.post("/api/events", async (c) => {
  const { userId, event } = await c.req.json();
  const result = await agent.respond([
    { role: "user", content: `Update profile for ${userId}: ${event}` },
  ]);
  return c.json({ userId, response: result.output });
});

app.get("/api/recommendations/:userId", async (c) => {
  const result = await agent.respond([
    { role: "user", content: `Recommend products for ${c.req.param("userId")}` },
  ]);
  return c.json(result.output);
});

No special runtime, no sidecar — it's a regular TypeScript import. Deploy it however you deploy your service.

Why

Agent skills work great inside local coding agents. Deploying them to production means rewriting everything.

Quick Start

Try an existing demo

git clone https://github.com/yujiachen-y/agent-bundle.git
cd agent-bundle/demo/personalized-recommend
npm run setup

Start your own project

npm install agent-bundle          # add to your project
npx agent-bundle generate         # generate typed client from agent-bundle.yaml

Then import and use in your code:

import { PersonalizedRecommend } from "@agent-bundle/personalized-recommend";

See Configuration Guide for all YAML options and Agent Skills for the skill format.

Features

Type-safe codegen

Prisma-style generate command. Your YAML config becomes a typed TypeScript package in node_modules/@agent-bundle/<name> — import it like any other dependency. When you use prompt variables, their names are checked at compile time.

Dev-prod parity

dev, serve, and build share the same sandbox abstraction. What passes locally ships as-is. No environment-specific surprises.

No vendor lock-in

Swap model providers or sandbox backends with one line of YAML. Run locally with Ollama — no API key needed:

model:
  provider: ollama
  model: llama3

Or use any cloud provider:

model:
  provider: anthropic            # or: openai, gemini, openrouter
  model: claude-sonnet-4-20250514

Plugins — turn any plugin into an agent

Pull skills and commands from a plugin marketplace and compose them into a single agent:

skills:
  - path: ./skills/report-formatter

commands:
  - path: ./commands/quick-analysis

plugins:
  - marketplace: anthropics/knowledge-work-plugins
    name: finance
    skills:
      - variance-analysis
      - close-management

Local skills, local commands, and marketplace plugins — all merged into one typed bundle. See the financial-plugin demo.

Live sandbox view

npx agent-bundle dev opens a WebUI at localhost:3000 — watch the agent's file tree, preview generated files, inspect the full LLM transcript, and monitor token usage and tool call metrics. No more black boxes.

Workspace	File Preview	Transcript	Metrics

OpenTelemetry tracing and metrics are also supported for production deployments.

Session recovery

Agent crashes mid-run? Resume from its last conversation state.

Token-scoped MCP

Connect to internal services via MCP servers. Even under prompt injection, the agent cannot exceed what the MCP server permits for that token.

Deploy

AWS ECS Fargate (beta)

npx agent-bundle deploy --target aws --secret API_KEY

Pushes to ECR and deploys to ECS Fargate — no Terraform or CloudFormation required. See Deploy docs for details.

Architecture

The agent orchestrator routes between the LLM provider, sandbox, and MCP servers. All three interfaces share the same abstraction across dev, serve, and build modes.

Demos

If agent-bundle is useful to you, consider giving it a star. It helps others discover the project.

Roadmap

• deploy --target aws GA — currently beta, stability not guaranteed
• GCP Cloud Run deploy target
• Pluggable agent loop engines — Claude Code, Codex via process bridge
• Fine-grained Docker sandbox isolation

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

License

MIT