Bring the skill.
BaoBox runs the agent.

The production runtime for AI agents. Tool calls, memory, guardrails, tracing and evals are already built — so your next agent is a new skill, not a new project.

The BaoBox bao opening its box: inside are the skill's system prompt, its tools and its model settings
  1. Guardrails Checks before and after the model Passed
  2. Tracing Every step, in order, with timings trace 0182a7 · 412 ms
  3. Evaluation Your test cases, scored on a schedule 12 of 12 cases passed
  4. Delivery API · Telegram · Email · Portal POST /api/v1/chat → 200 OK

Invite-only for now. We'll email you when there's room.

Example run with illustrative data.

What you don't have to build again

Every agent needs the same plumbing before it can do any real work. In BaoBox it's already there, and it's the same for every skill you add.

  • Tool calling. The loop that lets the model call your tools and use the results.
  • Sessions and memory. Multi-turn conversations that keep their history.
  • Keys and permissions. API keys that can only do what you allow, with rate limits.
  • Secrets. Credentials for your tools, encrypted and never shown back.
  • Guardrails. Checks on the request and on the answer.
  • Tracing. Every model call and tool call, recorded for every run.
  • Evals. Test cases with the answers you expect, run on a schedule.
  • Channels. The same skill on the API, Telegram, email and a portal.

What goes in the box

A skill is the part only you can write: what the agent is for, the tools it may use and the model it runs on. BaoBox runs it the same way every time, on every surface.

skill: support-agent
prompt: |
  You help customers with orders.
  Answer from the policy docs only.
tools:
  - lookup_order
  - create_ticket
model: your default, or your own key
guardrails:
  before: check the request
  after:  check the answer
Prompt
What the agent is for, in plain words.
Tools
The actions it can take — look something up, file a ticket, call your service.
Model
Use the platform's model or bring your own key. Swap it without rewriting the skill.
Guardrails
Checks before the model sees a request and after it answers.

Delivered where the work happens

The same skill, the same rules, wherever your users and your team already are.

  1. API

    Call it from your product: chat, streaming, or structured workflow output.

    POST /api/v1/chat
    { "skillId": "support-agent",
      "message": "Where is order 4417?" }
    
    200 OK
    { "response": "It shipped on Monday…" }
  2. Telegram

    Each team gets its own bot, with topics routed to the right skill.

    Where is order 4417?

    It shipped on Monday and arrives Thursday.

  3. Email

    Send through your own provider, from your own address.

    To: a customerYour order has shipped

    Hi — order 4417 left our warehouse on Monday and should arrive Thursday.

  4. Portal

    Your team signs in to manage skills, keys, models and runs.

    RunSurfaceGuardrails

    0182a7APIpassed

    0182a6Telegrampassed

Illustrative examples.

Every run gets a receipt

When an agent answers a customer, you should be able to see exactly what happened — and whether it was good.

  • Trace every step. Guardrails, model calls and tool calls, in order, with timings.
  • Score it before you ship. Write eval cases with the answers you expect, run them on a schedule, and compare one run with the last.
  • Guard both ends. Check the request before the model sees it, and the answer before anyone else does.
Run receipt 0182a7
  1. Guardrail · beforepassed
  2. Model call980 tokens · 640 ms
  3. Tool · lookup_order120 ms
  4. Guardrail · afterpassed
  5. Answered · API1.1 s total

ResultPassed

Illustrative data.

Your model, your keys

Start on the platform's model, or connect your own keys for OpenAI, Anthropic, Gemini, MiniMax or Workers AI. Pin a model per skill, and fall back to another provider when one is down.

A box for every client

Building agents for several businesses? Each client gets its own skills, portal, keys and data, kept apart from everyone else's. API keys carry only the permissions you give them, and stored credentials are encrypted.

Pricing arrives with launch

We're opening BaoBox to a small number of teams first. Join the waitlist and you'll hear about pricing before anyone else.

Join the waitlist

Questions

Can I use BaoBox today?

BaoBox is invite-only while we onboard the first teams. Join the waitlist and we'll be in touch.

Which models does it support?

OpenAI, Anthropic, Gemini, MiniMax and Cloudflare Workers AI. Use the platform's key or bring your own — Gemini works with your own key.

Is my data kept separate from other teams?

Yes. Every team's skills, runs and keys are isolated, and the credentials you store are encrypted.

Can I host it myself?

Not today. BaoBox runs as a hosted platform.

How much does it cost?

Pricing will be announced at launch. People on the waitlist hear first.

There's room in the box for your team

We only use your email to tell you about BaoBox.