Bring the AI you already use
Any AI that speaks MCP — or can call an OpenAPI tool — can work in Zeze under your name. Security doesn’t depend on which AI you pick: every action is checked on our server.
Connected in three steps
Copy your address
In the app, open Me → AI assistant permissions → Authorized AIs and copy your workspace’s MCP address.
Paste it into your AI and sign in
Your AI opens a Zeze sign-in page in the browser. You choose what it may do — read contacts, read messages, send messages, manage groups.
Your AI works as you
It reads, drafts and follows up in your name. Pause or revoke it any time from the same page; revoking takes effect immediately.
Works with
Setup steps follow each vendor’s documentation as of October 2026. AI vendors change their settings often — the in-app guide is kept current.
| AI | How it connects | Status |
|---|---|---|
| Claude (web, desktop, mobile) | Settings → Connectors → Add custom connector → paste address → sign in | Supported |
| Claude Code | claude mcp add --transport http, then /mcp to sign in | Supported |
| ChatGPT | Developer mode connector with OAuth (your ChatGPT workspace must allow custom connectors) | Supported |
| Custom GPTs | GPT Actions: import our OpenAPI schema, OAuth with a client ID/secret generated in Zeze | Via OpenAPI |
| OpenAI API / Agents SDK | MCP tool with a token obtained by your app’s OAuth flow | Supported |
| Gemini CLI | httpUrl + oauth in settings.json, then /mcp auth | Supported |
| Gemini web app | No personal custom-MCP option found yet — use Gemini CLI or another MCP client | Not yet |
| Cursor | Add the address to mcp.json, then Login in settings | Supported |
| VS Code + GitHub Copilot (agent mode) | Add an http server to mcp.json and follow the sign-in prompt | Supported |
| Kimi (Kimi Code CLI) | kimi mcp add --transport http --auth oauth | Supported |
| Tongyi / Alibaba Cloud Model Studio | Add a custom MCP service and bind it to your agent app; if OAuth isn’t available there, an admin generates an OpenAPI client | Supported |
| Doubao / Coze | Create a plugin from our OpenAPI schema with OAuth authorization code | Via OpenAPI |
| DeepSeek | Use an MCP-capable client (e.g. Cherry Studio) with DeepSeek models | Via an MCP client |
| Zhipu GLM | Use an MCP-capable client with GLM models | Via an MCP client |
| Any other MCP client | Paste the address; for clients without OAuth, use a bridge such as mcp-remote | Supported |
You choose what it can do
Permissions are picked on the sign-in page and can be narrowed later. Your admin sets the ceiling — if they lower it, connected AIs lose those tools immediately.
| Permission | What it allows |
|---|---|
| Read contacts | Look up colleagues and your own profile |
| Read messages | List conversations, read and search messages, unread overview, wait for replies |
| Send messages | Send messages and files as you — sensitive ones wait for your approval |
| Manage groups | Create groups and invite people as you — sensitive ones wait for your approval |
Default rules — enforced on the server
Each AI action is Auto, Ask me, or Never. You can tighten any rule; admins can set minimums that individuals can’t lower. When several rules match, the strictest wins.
| Action | Default |
|---|---|
| Direct message to a colleague | Auto |
| Post in a group up to 20 people | Auto |
| Post in a larger group | Ask me |
| Message an important contact (your list: manager, clients…) | Ask me |
| Anything involving another company or an outside account minimum — can’t be set lower | Ask me |
| A link to an outside domain minimum — can’t be set lower | Ask me |
| @everyone | Ask me |
| Recall your own message | Ask me |
| Delete someone else’s message | Never |
| Change group settings, kick, leave | Never |
| Change AI rules, devices, password, push settings always blocked | Never |
Security model
The server decides
All checks run in your workspace’s gateway, not in the AI, the plugin or the app. Switch AIs or call the raw API directly — the same rules apply.
Your AI never holds your credentials
It gets a short-lived token issued by the gateway (stored only as a hash). Your messaging credentials stay encrypted (AES-256-GCM) inside the gateway.
One AI, one device
Each authorization is its own device, shown by name. Pause or revoke one without touching the others; the audit log records which AI did what.
Other people’s words are data, not instructions
Messages, names and group titles are wrapped in randomized “untrusted data” boundaries before reaching the AI. Hidden characters are stripped; fake system markers are neutralized; outside senders are labeled.
No quiet exfiltration
Auto-loading images and embeds are stripped from AI-sent content. Outside links and outside recipients always need your approval. Heavy reading followed by an outbound send pauses the AI automatically.
Admins hold the master switch
Turn AI off for the whole workspace, allow only verified AI clients (identified by their redirect address, not their self-reported name), cap permissions and quotas.
Everything is recorded
Every AI action is logged with time, person, device, AI client, action, target, decision, matched rules and which conversations it read. Authorizations, revocations, rate-limit hits and anomaly pauses are logged too.
Rate limits and quotas
Defaults: 10 write actions per minute and 300 per day per AI, 5,000 messages read per day. Admins can adjust them.
How this differs from a chat bot integration
| Typical bot or assistant integration | Zeze | |
|---|---|---|
| Who is speaking | A bot account, or a built-in assistant speaking as itself | You — through your AI, labeled “via Claude”, “via ChatGPT”… |
| What it can see | Channels the bot was added to | Exactly what you can see, nothing more |
| Who decides what needs approval | Usually the AI client, or nobody | The server, using your rules and your admin’s minimums |
| What recipients see | A bot message | “via AI” or “via AI · confirmed by you”, set only by the server |
| Turning it off | Remove the app | Pause or revoke one AI instantly; admins can switch AI off workspace-wide |
Coming soon
- Approve or discard drafts from a phone notification
- A 30-second undo window after approving
- Draft bar above the message box, Tab to accept
- Limit an AI to specific conversations
- AI meeting notes and action items after calls
- Built-in AI features with model routing and BYOK