Build companies that run on AI agents. Not companies that use AI -- companies where agents ARE the employees.

npx @abf/cli init --template solo-founder --name my-business
cd my-business
abf auth anthropic
abf dev
# Open http://localhost:3000

Three agents are now running: an executive assistant, a researcher, and a content writer -- coordinating through a message bus, sharing memory, and reporting to you through a web dashboard. That took under 2 minutes.

Most AI agent frameworks are developer tools. ABF is a business-building framework. The difference:

• Describe your business, get an agent team. ABF's Seed-to-Company pipeline reads your business plan (or interviews you about your idea) and generates a complete agent team -- roles, tools, knowledge base, workflows, and project structure. No configuration required.
• Two users, one framework. Operators manage agents through a visual Dashboard. Builders work in YAML, TypeScript, and the CLI. Both see the same system.
• Files are the API. Every agent definition, memory file, and configuration is a plain file (YAML, Markdown, JSON). Version control, code review, and diffing work out of the box.
• Security is structural, not optional. Behavioral bounds are enforced by the runtime, not the LLM. Agents cannot bypass their permissions. Approval queues gate sensitive actions. Everything is audited.

ABF ships with 3 templates and can generate custom teams from any business description:

Have a business plan? Generate a custom team -- and a plan to build the product -- in one command:

abf init --seed ./my-business-plan.md

ABF accepts .docx, .pdf, .txt, and .md files. When your seed document describes a product to build (SaaS, marketplace, platform), the analyzer generates an adaptive build plan alongside the agent team. A Builder agent orchestrates product construction -- provisioning databases, generating frontends, configuring payments, deploying -- with human approval at every critical step. Full Seed-to-Company guide.

ABF is built on 6 primitives: Agents (autonomous workers with roles and tools), Teams (agent groups with an orchestrator), Memory (5-layer persistence from session to knowledge base), a Message Bus (typed inter-agent communication), Tools (30+ built-in, MCP servers, and custom .tool.js), and Triggers (cron, event, message, webhook).

The runtime is a single Node.js process with five components:

Scheduler fires triggers. Dispatcher spawns work sessions. Session Manager runs an 8-step lifecycle: load context, build prompt, call LLM, execute tools, route messages, write memory, check escalations, log results. Bus routes inter-agent messages. Gateway serves the REST API, SSE events, webhooks, and Dashboard on a single port.

Read the Concepts Guide for the full explanation.

Autonomous agents run with real credentials and take real actions. ABF is built for containment:

• Behavioral bounds enforced by the runtime (not the LLM) gate every action
• Credential isolation -- each tool only sees the API keys it needs (ScopedVault)
• Execution sandboxing -- code runs without access to credentials or home directory
• Approval queues -- sensitive actions require human sign-off before executing
• Prompt injection defense -- source tagging, content isolation, and detection pipeline
• Full audit trail -- every session, tool call, message, and memory write is logged

No framework makes autonomous agents completely safe. The Security Guide is transparent about what ABF protects against, what it cannot, and how to configure your deployment for safety. Required reading before going to production.

abf deploy --target railway    # or: render, fly

Docker, Railway, Render, and Fly.io are all supported. See the Self-Hosting Guide for production configuration including PostgreSQL, Redis, TLS, and API authentication.

git clone https://github.com/alexclowe/abf.git
cd abf
pnpm install    # requires pnpm 10+
pnpm build
pnpm test

The monorepo has three packages: packages/core (runtime, providers, tools), packages/cli (CLI and templates), and packages/dashboard (Next.js dashboard). See CONTRIBUTING.md for conventions and guidelines.

MIT