The new wizard

Answer a few questions and bxAgents new builds the agent: tools, MCP servers, channels, access and deploy target.

The new wizard

Run bxAgents new in a terminal with no --model and it asks what the agent should do, then writes the project to match. Nothing in it is magic: every answer ends up in a file you can read, and the last screen prints the one-line command that would build the same project without any questions.

bxAgents new my-agent

It is a full-screen UI in the teal of the BxAgents logo. Arrow keys (or j/k) move, Space ticks, Enter confirms, Esc goes back one question, Ctrl+C cancels and creates nothing. Answers you have given collapse to one line each above the current question.

What it asks

A stage whose answer you already gave as a flag is skipped and shown as answered. A stage that does not apply (the search provider when search is off) is skipped without a word.

File tools never get the whole disk

bx-ai ships 19 file tools but registers none of them. The wizard offers them, and always asks which folders they may use: the temp folder (tmp), the folder you run the agent from (cwd), or a path you type. It will not continue with none. There is no way to ask for unrestricted access from the wizard or from the flags; an empty list is refused because bx-ai reads an empty list as "everywhere".

tools:
  - readFile@bxai
  - listDirectory@bxai
toolOptions:
  bxai:
    allowedPaths:
      - tmp
      - cwd

Secrets

The wizard never asks you to type a secret that it stores.

• API keys and tokens go in .env as empty stubs, each under a comment saying where to get it. The files that use them hold only the variable's name.
• A web access key is generated for you (a random 64-character value) and written to .env, because it protects something you are about to run.
• A web chat password is hashed with the same PBKDF2 hasher as bxAgents hash-password; only the hash reaches gateways/webui.bx, and it is safe to commit.

.env is in the .gitignore the project is created with.

MCP servers

All servers in the picker are remote (HTTP) servers that the bundled bx-ai can connect to, listed with how risky they are. Each starts with its credential read from the environment when the app starts:

mcpServers:
  - name: github
    url: "https://api.githubcopilot.com/mcp/"
    token: "${env.GITHUB_PERSONAL_ACCESS_TOKEN}"

Picking a high-risk server shows what it can do before you continue. Git is in the catalog but not offered yet: it runs as a local program, and the bx-ai bundled with this release talks to web MCP servers only.

After it is created

new runs box install (skip it with --skipInstall), then prints what is left: the keys to fill in, how to chat, test, serve and deploy, and, for each channel you chose, the setup that happens outside the project. For WhatsApp that means the Meta app, the verify token and a public address; for Slack, Socket Mode and two tokens; for Signal, a running signal-cli. The same steps live on each gateway page.

Without the wizard

The wizard writes nothing a flag cannot. The last screen prints the command; this one came from a real run:

bxAgents new support-bot --name=support-bot --model=claude --with=time,fetch,search --search=brave \
  --mcp=github --channels=web,telegram --web-auth=key --guard=flag --deploy=docker --registry=ghcr:me/support-bot

A flag with a bad value stops the command and says what the flag accepts. The wizard turns on when:

Set NO_COLOR for plain output. The Agent.yaml options the wizard writes need the default Agent.yaml, so they cannot be combined with --class.