Ask, understand,
把捕获的流量变成答案，以及受角色约束的操作。
- RBAC 范围控制
- 需确认
- 密钥已脱敏
有据可查，从不瞎猜
问一句“这个应用是怎么登录的？”，就能直接从你的抓包中拿到真实流程，外加可运行的复现代码（密钥已脱敏）。每个答案都带证据。
在哪里都记得你
教它一次，它就在每个界面跟随你——仪表盘、Telegram、MCP 和终端。想删掉什么，一句“忘了它”即可。
不会泄密的团队大脑
团队知识从所有人的对话中提炼，经过严格的隐私过滤和人工审核队列。个人事实与团队存储在架构上彻底隔离——设计上就无法串流。
随叫随到的专家
把调校好的人设绑定到某个 API 的服务组，只要那些流量成为焦点，BusyBro 就会自动切换。支付专家在支付流量出现时登场——无需任何人开口。
可验证的诚实
活动轨迹精确展示每个答案背后运行了哪些工具，BusyBro 的每次写入都会进入平台审计日志。信任它，因为你随时可以验证。
用什么模型，你说了算
运行默认的 Anthropic 模型，指向任意 OpenAI 兼容的端点，或接入本地模型。密钥存在加密保险库里，记忆在切换后依然保留。
Ask once,
One brain behind Telegram, the iOS app, the terminal, and your MCP-connected agents — same account, same memory, same rules.
Link once
01On Telegram, /login pairs the bot with your account via a short-lived code — no link, no LLM call at all. The iOS app and the bmc terminal already speak as you.
Ask in plain language
02Name a device by its name, PAC port, or URL; name a host; describe a bug. Attach a screenshot or a PDF on Telegram and it reads those too.
Tools, not vibes
03Answers come from real tool calls over your captured data — drawn from the same registry the dashboard and the MCP server consume — never from guessing at your traffic.
You hold the pen
04Reads flow at your role. Anything that changes state stops and shows you the exact payload first — an inline Confirm on Telegram, an action card elsewhere.
Autonomy with a
Every gate is re-checked at execution time, server-side — the model can neither grant itself a role nor press its own Confirm button.
| You ask it to… | What actually happens | Enforced by |
|---|---|---|
| Read traffic | Owner-scoped by default: it inspects your own devices' captures. Reading wider requires the devices-view capability on your real role. | Row-level security plus your role — the same checks as the dashboard. |
| Change something | It stops and shows the exact change — the script body, the settings diff, the delete scope — and waits for your explicit approval. | A single-use, server-side confirmation the model cannot forge or replay. |
| Do something admin-grade | Admin-only tools re-resolve your current role at the moment of execution — not when you asked, and not from the chat context. | Capability and admin gates re-applied at run time, defense in depth. |
| Touch the database directly | It can't. Raw database and broadcast primitives are excluded from the free-text model entirely — no dashboard screen grants raw SQL, so neither does your mate. | A code-level denylist between the model and the tool registry. |
| Handle a secret | Captured tokens and cookies are treated as secrets in its answers, and provider API keys are referenced by name only — the value never sits in chat-readable settings. | The encrypted vault behind one audited decrypt boundary. |
| Anything at all | Every write, every denial, and every error is recorded with the surface, your role, and whether a real human tap approved it. | The platform audit trail — the same one the dashboard tails live. |
Things people
Each ask maps to a real, bounded tool — not a fuzzy summary of rows it never read.
“How does this app's login flow work?”
It deep-inspects your captured traffic for that host — the auth calls, token refresh, required headers, payload shapes — and hands back repro code with credentials treated as secrets.
“What has my iPhone been doing for the last hour?”
A bounded aggregation answers with top hosts and request counts — never a raw scan of your rows — then drills into the noisy host when you ask it to.
“Clear the noise from that analytics host”
A targeted, filtered delete — and because it mutates your data, it first shows the exact scope it would remove and waits for your confirmation.
The questions sceptics
Can it act without asking me?
Reads run freely at your role. Every write stops for an explicit confirmation — and even if you auto-approve a tool for a thread, destructive and admin-grade tools always still ask.
Does it see my whole database?
No. Raw database primitives are excluded from the model entirely, and every tool it can call runs through the same row-level security and role checks as the dashboard.
What does it remember, and can I erase it?
Personal memory is owner-scoped rows you can list, search, and delete — say 'forget that' or send /forget on Telegram to wipe it. The background memory writer can be switched off for your account.
Can my private details leak into the shared team brain?
The team store is a separate table with no user column at all, a deterministic code filter rejects names, emails, IPs, paths and token-shaped strings before anything is stored, and every candidate waits in a review queue before it becomes recallable.
Which model answers — and can we run our own?
Anthropic Claude by default, configurable per surface. Admins can point the fleet at any OpenAI-compatible endpoint or a local model — the key stays in the encrypted vault — and on non-frontier providers the tool surface automatically narrows to a read-only core.
Is there a record of what it did?
Yes — every write, denial, and error lands in the platform audit trail with the surface it came from, the role that ran it, and whether a real human tap confirmed it.