ax is an open source framework for building AI agents that remember context, adapt to your workflows, and actually get better over time. We're figuring out agentic systems together — come build with us.
Not another chatbot wrapper. ax agents learn your context, use tools, work across models, and keep getting better at the things your team does every day.
Your agent remembers what matters. Semantic memory with embedding search, LLM-powered extraction, and proactive context recall — so you don't have to re-explain everything every session.
Same config on your laptop or across a Kubernetes cluster. Local sandboxing for personal use, K8s pods for your whole team. Scale when you're ready, not before.
Organize models by task — fast, thinking, coding — each with fallback chains. The router picks the right model for each job so you get speed where it matters and power where you need it.
Extend ax with community-built providers via the Provider SDK. Give your agent new skills, connect new services, add new models. If you can write TypeScript, you can build a plugin.
Agents that work while you don't. Scheduled tasks, proactive hints, and a heartbeat system that keeps things running. Set it up on Friday, find the results on Monday.
Watch your agent think. SSE event streaming, OpenTelemetry tracing, Langfuse integration. When something goes sideways, you'll know exactly what happened and why.
Drop-in OpenAI-compatible API, Slack integration, MCP servers, and channels for wherever your team already works. Point your existing tools at ax and they just work.
15 provider categories — LLM, memory, channel, database, sandbox, and more. Every piece is a TypeScript interface. Don't like something? Swap it. Want something new? Build it.
Sandboxed execution, taint tracking, prompt injection scanning, credential isolation. The guardrails are there so you can let your agents loose without losing sleep.
Same framework, same config. Start with a personal agent on your laptop. When the rest of your team wants in, scale to Kubernetes without rewriting anything.
# Your personal agent
profile: standard
models:
default:
- anthropic/claude-sonnet-4-20250514
providers:
memory: cortex
sandbox: seatbelt
skills: gitnpx ax init# Same agent, now for the whole team
replicaCount: 3
config:
models:
default:
- anthropic/claude-sonnet-4-20250514
- groq/llama-3.3-70b-versatile
providers:
memory: cortex
database: postgresql
sandbox: k8s
eventbus: postgres
audit: database
postgresql:
enabled: trueax agents don't just autocomplete — they break down problems, use tools, check their work, and iterate. Extended thinking models stream their reasoning in real time. It's weirdly satisfying to watch.
Every decision your agent makes is logged and queryable. Stream events in real time, plug into OpenTelemetry, or just tail the logs. When something goes weird — and it will, this is AI — you'll know exactly where to look.
Every piece of ax is a TypeScript interface. Swap Anthropic for OpenAI, SQLite for PostgreSQL, local process for Kubernetes pods. Build your own provider, publish it as a plugin, share it with the community. That's kind of the whole point.
Not every task needs your most expensive model. Organize by role — default, fast, thinking, coding — each with its own fallback chain. The router handles failover automatically. Your API bill will thank you.
Seriously, that's not marketing. It's three commands.
# Install and run
npm install
export ANTHROPIC_API_KEY=your-key-here
npm start
# Or use the CLI
ax configure # walks you through setup
ax serve # start the server
ax chat # talk to your agent
ax plugin add @ax/web # give it new skillsax is MIT-licensed and evolving fast. We're exploring the frontier of agentic systems in the open — new provider categories, better memory, smarter routing. If that sounds fun, jump in.