Connectors

A connector links your organization to a remote MCP server — Google Drive, a ticketing system, your own internal API — so your agents can call the tools that server offers. You add and manage connectors yourself in the console under Connectors.

Writing your own MCP server? See Build an MCP server for the authentication standards and tool design Harnix expects.

Name each connectionCopy Anchor Link

A connector has a name you choose, unique in your organization. Two connectors can point at the same server with different accounts — "Google Drive – HR" and "Google Drive – Finance" — and be given to different agents.

Add a connectorCopy Anchor Link

1. Open Connectors and click Add connector.
2. Enter a Name and the server's Remote MCP server URL.
3. Pick an Authentication method and click Add connector.

Harnix saves the connector and checks the server right away (up to 10 seconds). The sheet then shows one of:

• Connected — Found N tools. Done.
• Needs sign-in. Click Connect (see below).
• Error, with the reason. Fix the settings (often under Advanced) and click Save and retry. The connector is already saved either way.

AuthenticationCopy Anchor Link

Under Advanced you can set your own OAuth Client ID and OAuth Client Secret (only if the server doesn't support automatic registration, or you want to use your own OAuth app), the Scopes to ask for, and the Transport (Automatic tries Streamable HTTP first, then SSE).

If you create your own OAuth app with the provider, register this redirect URI: https://app.harnix.vn/connectors/callback.

Sign in (Connect)Copy Anchor Link

For an OAuth connector, Connect opens the provider's sign-in in a small window. Approve access there; the window closes itself and the connector shows Connected with its tools. If your browser blocks the window, click Continue in this tab.

The sign-in belongs to the organization: every conversation with an agent that uses this connector acts with the access you granted. Use separate connectors (and separate provider accounts) when different agents should see different data.

Choose tools per agentCopy Anchor Link

Open an agent, then its Tools tab:

• Knowledge search shows whether the agent searches its knowledge bases (see Tools).
• Built-in tools manages platform tools like ask_user that let the agent ask the user questions (see Tools).
• Connectors lists the connectors this agent may use. Add connector lets you pick one and choose its tools. Every tool is on by default — including tools the server adds later — and each has its own switch.

A tool you turn off is never offered to the model for this agent. Other agents using the same connector are not affected. Changes apply to runs that start after you make them.

Confirmation before a tool callCopy Anchor Link

A connector tool may change data in your systems — create an order, send an email, update a ticket. Such tools never run silently: the agent pauses, the end-user sees the connector's name, the tool's name and the real arguments the agent is about to send, and only then approves or declines. No write happens that the user has not seen.

Which tools need confirmation comes from the MCP server's annotations: a tool marked readOnlyHint: true (read-only) runs right away; every other tool — including those of a server that declares no annotations at all — asks first. Open an agent, go to its Tools tab, under Connectors: each tool has its own Needs approval switch to require or skip confirmation. The switch shows the default if there's no override; click Reset to default to remove an override. A tool you turn off (disabled_tools) cannot be changed — the switch is disabled.

In a conversation the confirmation goes through the approval frame of the stream and POST /v2/runs/{run_id}/approval: approve, approve for the whole conversation, or decline; the server-to-server channel decides for itself through the same route. Declining, or not answering within 5 minutes, doesn't break the conversation: the tool gets an error result, the assistant tells the user the action was not carried out, and the run ends normally.

Every decision — who, when, outcome, whether it was remembered for the conversation — shows on the run's trace in Runs & traces.

When a connector breaksCopy Anchor Link

A broken connector never makes a chat fail: the agent answers without that connector's tools until you fix it, and the run's trace records why.

Deactivate (in the connector's menu, or on its page) turns a connector off without losing its settings or sign-in: agents keep it assigned but can't call its tools until you click Activate.

If the Tools column shows —, Harnix hasn't listed that server's tools yet; click Test (or Refresh on the connector's page) to load them.

Disconnect clears the stored sign-in (the agents keep the connector assigned, its tools stop working until someone connects again). Delete removes the connector, its sign-in and its assignment to every agent.

Limits and securityCopy Anchor Link

• Only public https addresses are allowed; private or local addresses and URLs with a username or password are refused.
• Tokens, client secrets and header values are encrypted at rest and never shown again — not in the console, not in a trace.
• A single tool result is capped at 64 KB (see Tools).