Skip to content

MCP

The Model Context Protocol (MCP) is an open protocol, created by Anthropic, that connects AI agents to tools and data. Orqane supports MCP in both directions: your employees use MCP servers, and each employee can be an MCP server that other AI tools call.

An MCP server exposes three kinds of things:

MCP primitive What it is How Orqane uses it
Tools Actions the agent can call, with a name, a description and an input schema Become tools you grant to employees
Resources Read-only data, such as files or records Become searchable sources for the employee, like a knowledge base
Prompts Reusable instructions offered by the server Can be imported as the starting point of a skill

Orqane connects to servers over the two standard transports: Streamable HTTP for remote servers, and stdio for local servers, which Orqane runs for you in an isolated container.

Open the MCP settings

Go to Settings → Capabilities → MCP servers and click Add server.

Enter the server URL

Paste the server’s Streamable HTTP URL, for example https://mcp.linear.app/mcp, and give the server a name.

Authenticate

Choose how Orqane signs in:

Method When to use it
OAuth The server supports MCP authorization. Orqane opens the provider’s sign-in page and stores the tokens. Recommended.
API key The server expects a static key. Enter the header name, such as Authorization, and its value.
None Public servers only.

Choose Workspace credentials to let all granted employees share one account, or Per member to have each employee act with its human manager’s account.

Review the discovered tools

Orqane lists every tool, resource and prompt the server offers. Click Save.

Orqane gives each discovered tool a default impact class, based on its name, description and MCP annotations such as readOnlyHint and destructiveHint. A Builder reviews and edits the class on the server’s Tools list.

Tool Default class
list_issues Read
create_issue Write internal
post_comment on a public repository Communicate
delete_project Irreversible

A new tool is disabled until a Builder reviews it. If the server later adds tools or changes a tool’s description, Orqane flags the change and keeps the new version disabled until it’s reviewed. This protects you from a server that silently changes what a tool does.

On the employee’s Tools tab, click Add tools, choose the server, and select individual tools. You can also add MCP tools to a skill, so every employee with that skill gets them. Grant only what the job needs: a support employee rarely needs delete_project.

Orqane checks each server every few minutes and on every run.

Status Meaning
Healthy The server answers and the credentials are valid.
Degraded Calls are slow or some fail. Runs continue and retry.
Auth expired The OAuth token can’t be refreshed. A member must reconnect.
Down The server doesn’t answer. Runs that need it pause at a checkpoint.

Every employee, and every team, can have its own MCP endpoint. Any MCP client can then give work to the employee as if it were a tool.

Enable inbound MCP

An Admin turns on Inbound MCP in Settings → Capabilities.

Publish the employee

On the employee’s profile, open Integrations → MCP and click Enable MCP endpoint. Orqane shows the endpoint URL, for example https://mcp.orqane.com/acme/ada.

Choose who can connect

Select OAuth to let workspace members sign in with their Orqane account, or create an MCP access token for a service account. Tokens are scoped to one employee or team and can expire.

Tool What it does
ask Asks the employee a question and returns its answer. Runs as a short task with Read and Draft actions only.
assign_task Creates a task with a goal, inputs and an optional deadline. Returns a task ID.
get_task_status Returns the task’s status, its latest steps and its result when done.
cancel_task Stops a task and its current run.

A team endpoint offers the same tools. Tasks go to the team’s manager employee, which delegates them.

In Claude, open Settings → Connectors, click Add custom connector, and paste the endpoint URL. Sign in with your Orqane account when asked.

A task that arrives over MCP is a trigger like any other. The employee keeps its autonomy level, its policies and its thresholds. If it needs an approval, the approval goes to its human manager, and get_task_status reports waiting_for_approval until a human answers. Usage is charged to the employee’s budget, and every call appears in Activity with the client’s name.