An AI agent works on its own: it runs tasks on a schedule and sees them through, rather than just answering in a chat. To let it work with your Elbuz catalog, orders and customers, you give it a connection address and an employee's key.
When you need this:
- The agent should check orders every morning — it reads the data with the key by itself.
- You're not sure your agent can connect — the table shows the method for each one.
- The agent mustn't change anything without you — the key only reads or waits for your “Accept”.
Where to start: find your agent in the table.
What an agent can do in Elbuz
- Search and count — products, customers, documents, price lists.
- Propose changes — you accept them in “Pulse”.
- Work on a schedule — it checks what has changed by itself.
- Work in a browser tab — open windows and read the screen.
How to connect your agent
As of November 2026, based on the vendors' help pages. These products change fast: if your screen looks different, check the agent's own help.
| Agent | How to connect | Next step |
|---|---|---|
| OpenClaw | in the settings file | settings block |
| Hermes (Nous Research) | MCP server in the settings file | settings block |
| Grok Bot | MCP server by address | steps and limits |
| Meta Muse | MCP isn't documented; a custom connector to API v2 | how to ask Muse |
| OpenAI dots | a dot uses supported plugins installed and enabled in your account; the help says nothing about your own MCP server | check the OpenAI help on dots |
| Claude Duet | uses the apps connected to your Claude account; the help says nothing about custom connectors | check whether the Elbuz connector shows up in Duet after connecting it in claude.ai |
| Kimi Claw | the plugins Kimi uses to connect MCP are not supported in Kimi Claw | if you need an agent like this — OpenClaw on your own computer |
| Cue (Manus) | the vendor doesn't document custom servers for Cue | Manus itself has Settings → Integrations → Custom MCP Servers; for Cue, check the Manus help |
| Gemini CLI, Antigravity | MCP server via a command or the settings file | command, settings block |
| Claude, ChatGPT, Codex, Cursor | MCP server by address and key, or a connector | connect an AI assistant |
Can your agent run commands and read pages by itself? Give it this phrase: it will read the instructions and help you connect.
| What to tell the agent | What it will do |
|---|---|
| “Read the instructions: https://elbuz.com/en/guide-connect-ai-assistant-mcp#dlia-asystenta — help me connect Elbuz over MCP and work with it by these rules.” | asks for your Elbuz address and walks you through the steps, or connects by itself |
What to prepare before connecting
An autonomous agent works when you're not around. So give it separate, limited access rather than your own.
- Add a separate employee for the agent and give their department only the rights it needs.
- Create an on their behalf: steps.
- To start with, leave “Read only” on. When the agent needs to fix things, clear the checkbox and set “Data changes” to “Ask”: every change will wait for your “Accept” in “Pulse”.
- Leave “Allowed IP addresses” empty if the agent runs in the cloud: it connects from its own servers, and an address list will block it.
- Enter the key in the agent's settings yourself, not in the chat: the chat is kept in the agent's history.
- Sign in without a key on the “Allow” page as the agent's employee: the key is created on behalf of whoever is signed in to the browser, and without the “Read only” checkbox.
1
Data changes2
Allowed IP addresses3
Read only- Data changes — in the “ask” mode a change waits for your Accept in Pulse.
- Allowed IP addresses — leave empty for a cloud agent.
- Read only — leave it on to start with.
How to connect an agent with a settings file
In the examples, <your-domain> is your Elbuz address, for example 123456789.elbuz.com. The key is passed in the Authorization.
The key is created on behalf of whoever is signed in to the browser, without the “Read only” checkbox, and an employee from “Management” can pick “Act on its own” on the consent screen. Sign in as the agent's employee and choose “Ask”.
OpenClaw
OpenClaw is an open-source agent that runs on your computer or server. Add this to ~/.openclaw/openclaw.json:
{
mcp: {
servers: {
elbuz: {
url: "https://<your-domain>/api/mcp",
transport: "streamable-http",
headers: {
Authorization: "Bearer ${ELBUZ_API_KEY}",
},
},
},
},
}
Put the key in the ELBUZ_API_KEY environment variable (a system setting a program reads values from, so the key isn't stored in the file). Sign-in without a key: instead of headers, set auth: "oauth" and run openclaw mcp login elbuz. OpenClaw runs JavaScript in the browser by default (browser.evaluateEnabled), so it can also use elbuz_tools. Help: MCP settings in OpenClaw.
Hermes
Hermes Agent by Nous Research runs on your computer or server. Add this to ~/.hermes/config.yaml:
mcp_servers:
elbuz:
url: "https://<your-domain>/api/mcp"
headers:
Authorization: "Bearer ${ELBUZ_API_KEY}"
Write the key to ~/.hermes/.env as the line ELBUZ_API_KEY=elbz_xxxxxxxx_… and run /reload-mcp in the session. Sign-in without a key: instead of headers, set auth: oauth and run hermes mcp login elbuz. The tools appear as mcp_elbuz_*. Help: MCP in Hermes.
Hermes can use elbuz_tools only under the conditions in its help: browser_exec is available to a session with the terminal toolset, and browser_cdp when a browser connection over CDP is set up.
Gemini CLI
gemini mcp add --transport http --header "Authorization: Bearer elbz_xxxxxxxx_…" elbuz https://<your-domain>/api/mcp
By default, the server is added for the current project. For the free tier and Google One, Gemini CLI has been replaced by Antigravity CLI (help).
Antigravity
Add this to ~/.gemini/config/mcp_config.json:
{
"mcpServers": {
"elbuz": {
"serverUrl": "https://<your-domain>/api/mcp",
"headers": {
"Authorization": "Bearer elbz_xxxxxxxx_…"
}
}
}
}
Without headers, Antigravity offers sign-in by itself if the server supports it (help).
How to connect Grok Bot
The bot adds an MCP server by address by itself. Tell it:
| What to tell the agent | What it will do |
|---|---|
| “Add the Elbuz MCP server at my address /api/mcp. I'll enter the access key myself in a secure form — don't ask for it in the chat.” | adds the server and asks for the address; if the key can't be entered in a form, it tells you so |
The official help doesn't describe how to pass the bot a header with a key, and it forbids pasting passwords into the chat. Check the sign-in method in the Grok Bot help. Grok.com works differently: grok.com/connectors → New Connector → Custom, and the server must be reachable from the internet (help).
How to connect Muse
Meta doesn't document connecting MCP servers in Muse, and the Muse browser doesn't run JavaScript on the page, so elbuz_tools won't work here. That leaves a custom connector: Muse builds one by itself if the service has an API.
| What to tell the agent | What it will do |
|---|---|
| “Create a custom connector to the Elbuz API v2 using the docs at https://elbuz.com/en/guide-api-v2. I'll pass the key through the credential store — don't ask for it in the chat.” | walks you through creating the connector |
Meta doesn't review custom connectors. You can manage existing connectors in Settings → Connectors (Meta help). The API is described in API v2: keys, permissions, format.
How to run an agent on a schedule or event
On a schedule — the simplest way. The agent's requests changes by itself, for example once an hour: search with the updated_since filter. Filtering and an example are in the API v2 article.
On an event — through an intermediate handler. An Elbuz webhook sends its own signed format, without an Authorization header. It can't trigger these directly:
- OpenClaw —
POST /hooks/agentwhenhooks.enabledis on; it needs anAuthorization: Bearerheader and amessagefield in the body; - Grok Bot routines — need an
Authorization: Bearerheader; - Grok automations — a signature in the same Standard Webhooks format, but Grok issues the secret, while Elbuz signs with its own.
You need a small program of your own that checks the Elbuz signature and calls the agent.
Scenarios
Morning order report
OpenClaw with a key created in the window with “Read only” on: sign-in without a key doesn't set this checkbox. Every morning, a routine counts new orders with query and sends you a summary. It can't change anything.
The agent writes product descriptions
Hermes with a key in “Ask” mode. The agent drafts descriptions, each edit arrives in “Pulse”, and you accept them all at once.
A cloud agent won't connect
A cloud agent is refused because of your key: its “Allowed IP addresses” field is filled in. Clear the field and limit the agent with the employee's rights.
What to do when the agent won't connect
| What you see | Why | What to do |
|---|---|---|
| “IP … is not allowed for this key” | the agent connects from an address not on the key's list | clear the field (preparation) |
| “Authorization: Bearer <Elbuz API v2 key> is required” | the key wasn't passed or isn't valid | check the header (example) |
| “read-only key” | the key has “Read only” on | clear the checkbox (preparation) |
| The agent says it's “done”, but the data hasn't changed | the change is waiting in “Pulse” | accept it (reviewing changes) |
| “rate limit exceeded” | more than 120 requests per minute per key | run the routine less often (schedule) |
FAQ
Which connection method should I choose for my agent?
Where the agent supports MCP, use the Elbuz MCP server: it gives data and actions without a browser. If there's no MCP but the agent can call an API, use API v2 with a key.
Can I give the agent the key in the chat?
Better not: the chat is kept in the agent's history. Enter the key in the agent's settings yourself.
Can the agent change data without me?
A key created in the window starts with “Read only” on, and in “Ask” mode every change waits for your “Accept” in “Pulse”. Sign-in without a key on the “Allow” page creates a key without this checkbox, so sign in as the agent's employee and choose “Ask”.
Can Elbuz start the agent by itself?
Not directly: an Elbuz webhook doesn't send the Authorization header and has its own body format. You need your own intermediate handler. The simpler way is a scheduled agent routine that requests changes by itself.
Who pays for the agent's work?
You pay the agent's vendor for its work under your subscription, if it's a paid one. Elbuz doesn't charge anything for the agent's actions through the key.
What if my agent isn't in the table?
Check the agent's help to see whether it supports a remote MCP server by address with a key. If it does, it connects the same way as OpenClaw or Hermes.
Where it is and who has access
- MCP address
- your Elbuz address plus
/api/mcp, for examplehttps://123456789.elbuz.com/api/mcp - Key
- Account → System settings → “Other settings” → “API and access tokens”
- Agent requests
- Contacts → AI Pulse, “Proposals” tab
- Access rights
- the agent gets the rights of the key's employee
- Prepare
- a separate employee with the rights you need
Related topics
- If your assistant is Claude, ChatGPT, Codex or Cursor — connect an AI assistant over MCP.
- If the agent works in a browser tab — WebMCP in Elbuz and AI helper in the browser.
- If you're building your own integration — API v2 and webhooks.
