Build AI agents by convention.
Describe an agent with files and folders. bxAgents build assembles a real, runnable ColdBox application from them, once, before anything boots.
$ install-bx-module bx-agentsAgent.yaml is the agent
A name, a model and any settings you need, in plain YAML. Need code? Use a class-based Agent.bx instead.
Conventions first
Convention over configuration for faster development.
Pluggable and extensible
Swap models, tools, memory and more with ease.
Powerful agents
Agents that reason, act, and collaborate effectively.
Production ready
Built for performance, reliability, and real-world applications.
From nothing to a chat in four commands.
Install the module, scaffold an agent, build it, and talk to it. Nothing else to configure.
install-bx-module bx-agentsbxAgents new my-agent \
--model=openai/gpt-5cd my-agent
bxAgents buildbxAgents chat
# or serve over HTTP
bxAgents serve --port=8080Real runs, recorded.
Every clip is a real session against a real served agent. The model is the mock provider the examples ship with, so each answer is "Mock response" and no model was called.
From nothing to a running agent
You write the agent. The folders do the wiring.
This is the whole simple-agent example from the repo: a YAML file, a prompt, one tool. Add a folder when you need it, and the build picks it up.
$ bxAgents new support-bot --model=openai/gpt-5 --template=webui-chat Created new BX Agents project [support-bot] at /work/support-bot $ bxAgents build Build succeeded: /work/support-bot/.build/app $ bxAgents inspect Agent: support-bot Model: OpenAI/gpt-5 Environment: development Manifest ver: 1.0.0 $ bxAgents serve --port=8080 Serving [/work/support-bot/.build/app] on http://127.0.0.1:8080 - bound to 127.0.0.1 only
Discovery happens once, not on every boot.
Most agent frameworks wire tools, skills, routes and schedules together at request time. BxAgents does the work in bxAgents build and ships an ordinary app.
Discover
Walk the convention folders that exist and have content.
Validate
Reject bad config before anything runs.
Manifest
Record what was found in manifest.json.
Generate
Write config, handlers, router and scheduler.
Run
Boot .build/app like any ColdBox app, or ship a .bxa.
Wired at request time
Convention scanning and file walking on every boot, and build work deferred into the request path.
Wired at build time
No scanning at startup. Starting the agent is booting a plain ColdBox application.
Twelve gateways. Hover one to inspect it.
Drop a file into gateways/ and your agent is on that channel. Each tile shows the type you set and the environment variables it reads, straight from the docs.
Hover, focus or tap a tile.
- A generated chat UI: conversation history, streaming replies and a mobile layout.
- Light and dark theming, themeable from a
gateways/entry. - No front-end build step of your own.
exposes: "agent"Mounts invoke and info routes under a path you choose, like /api/chat.
exposes: "webui"A streaming chat front end at a path like /chat.
exposes: "mcp"Any MCP client can connect at a path like /mcp.
One gateway file, a working HTTP API.
One small file, gateways/expose.bx, and the build mounts a REST API for your agent. The output on the left is a real capture from a served agent.
Loading the captured session...class { function configure() { return { exposes : "agent", path : "/api/chat" }; } }
Captured on 2026-10-16 from the http-gateway-agent example on BoxLang 1.16.0. It uses the mock model provider, which answers every input with "Mock response", so no model was called. The routes on the right are read from the /info response.
Pick a starter. See exactly what you get.
Generated by the real scaffolder for each template. Select a file to read it.
$ bxAgents new my-agent --model=openai/gpt-5 --template=minimal--repo=owner/repo and --forgebox=<slug>.A 20-lesson course, from install to a deployed agent.
Follow the guided path through the same conventions you just saw. No account needed: the course remembers where you stopped in your own browser.
.bxa ready to deploy.Fourteen verbs, one executable.
Tab completion for bash, fish and PowerShell ships with the module.
Built with BoxLang. A whole platform.
Need agents in production? Talk to the people who build BoxLang.
Ortus Solutions offers professional support, architecture reviews and custom development for BoxLang, including agent design, gateways, tools and deployment.