Use the Nexra MCP server

What the MCP server gives you

Nexra exposes one hosted MCP endpoint. Each application in your workspace owns one or more MCP servers: a reviewed, versioned list of tools, resources and prompts. The credential an agent presents decides which application, server and published version answer the request.

Endpoint and protocol

Authentication

Nexra runs its own OAuth 2.1 authorisation server for MCP. There are two ways to get an access token.

Interactive clients (OAuth sign-in)

Clients that support MCP authorisation, such as Claude, ChatGPT, VS Code, Cursor and the MCP Inspector, need only the server URL. On first use:

Approving creates a credential named OAuth: <client name> on the application. Revoke it on the application's MCP credentials page to disconnect the client immediately. Members who can manage MCP (owners, administrators, developers, security administrators and approvers) can approve any published tools. A member of a read access group who can't manage MCP can approve for themselves too; their grant covers the Ask company data tools only.

Approve only clients you started yourself. Dynamically registered clients are not verified by Nexra; the consent page shows the address you will be sent back to.

Server-to-server agents (client credentials)

For an agent with no person present, create a credential under MCP → Credentials and exchange it for an access token. The credential's prefix (mcp_…) is the client ID and its secret is the client secret:

curl -s https://api.nexra.amakoragroup.com/oauth/token \
  -u "$NEXRA_MCP_CLIENT_ID:$NEXRA_MCP_CLIENT_SECRET" \
  -d grant_type=client_credentials

The response is {"access_token": "...", "token_type": "Bearer", "expires_in": 3600, "scope": "..."}. Request a new token when it expires. The secret itself is not accepted on /mcp in production.

A client-credentials token has no person behind it, so it can't use the Ask company data tools.

Endpoints

Access tokens are RS256 JWTs for the audience https://api.nexra.amakoragroup.com/mcp. They carry the workspace, application, server and credential they were issued for, and stop working as soon as that credential is revoked. Redirect URIs must be HTTPS, an http loopback address (any port) or a private-use scheme such as cursor://. Replaying an old refresh token revokes the whole grant.

Scopes

Imported tools use the scopes you assign during review (the console suggests mcp:tool:<tool_name>).

Set up a server in the console

Disable turns a server off and revokes every credential issued for it.

Tools

Connect tools need the application to be linked to a project, and they only see that project. Trigger tools take the flow or integration ID, an idempotencyKey and a correlationId. They are idempotent: repeating a call with the same idempotencyKey returns the original execution with replayed: true. get_connect_run takes an executionId.

Tools that create webhooks, make live API calls, change credentials or change billing are defined but disabled, and are refused if called.

Resources (nexra://api-catalogue, nexra://openapi/{apiSlug}/{version}, nexra://docs/{slug}, nexra://errors and others) and prompts (generate_integration_plan, debug_failed_api_call, create_sandbox_test_flow, generate_webhook_receiver_code, explain_api_response) are listed with resources/list and prompts/list.

Workspace records tools

These tools give an agent the same view of your workspace records as Nexra Agent: obligations, the customer map, evidence matrices and statements.

The credential acts as the person behind it (the person who approved the OAuth sign-in, or who created the credential) with that person's current workspace role. Read tools return only what that person can see. Proposal tools change nothing: each one stores a proposal, and a different workspace member who manages records (an owner, admin, Connect operator or developer) approves it in the console before anything is saved.

A proposal answers with status, actionId, a plain summary and confirmationText, and a note that nothing has changed yet. The evidence matrix tool follows the fairness rules: no scores, ranks or recommendations, no protected characteristics, and nothing is ever rejected automatically.

Ask company data

Two read-only tools let a leader ask about the business's own data where they already work: Microsoft Teams with Copilot Studio, Microsoft 365 Copilot, Claude, ChatGPT or your own agent. They follow exactly the rules of Ask company data in Nexra Agent.

As you, and only you. Both tools run only for a person who signed in themselves through OAuth. What you can read depends on your read access group, so a workspace API key, an MCP credential secret or a client-credentials token is refused with a plain message, and nothing is read. Your sign-in must still be an active member of the workspace.

The console's rules.

company_ask returns the cited answer (answer, sources, calculations, gaps) and text: the same answer as plain text with its citations, ready to show in a chat. Pass an answer's id as followUpOf to ask a follow-up.

To use the tools, an admin adds company_ask and leadership_summary to the MCP server and publishes a new version (new servers have them). For Teams, see Use Nexra in Teams with Copilot Studio. The same two calls are available over REST; see MCP API.

Calling the server

MCP=https://api.nexra.amakoragroup.com/mcp

curl -i "$MCP" \
  -H "Authorization: Bearer $NEXRA_MCP_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-11-25","clientInfo":{"name":"my-agent","version":"1.0.0"},"capabilities":{}}}'

Copy the MCP-Session-Id response header into SESSION_ID. Send notifications/initialized, then tools/list, before calling a tool. Include the session and negotiated protocol headers on each request. The connector setup guide shows discovery and exact operation selection:

curl "$MCP" \
  -H "Authorization: Bearer $NEXRA_MCP_TOKEN" \
  -H "MCP-Session-Id: $SESSION_ID" \
  -H "MCP-Protocol-Version: 2025-11-25" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"trigger_approved_connect_flow","arguments":{"flowId":"<flow-id>","idempotencyKey":"order-1042-sync","correlationId":"agent-run-1042"}}}'

Tool results are returned as one text content item that contains JSON.

Approvals for sensitive tools

Tools marked requiresConfirmation (including imported write operations) do not run on the first call:

Connect trigger tools are gated differently: they only run flows whose production deployment is already approved, for requesters who are still workspace members.

Limits, errors and logs

Current limitations

Next: Export MCP and API configuration.

Connect a provider account across interfaces

See Connect through MCP, APIs, CLI and SDKs for account setup, exact operation selection, authentication, copyable examples, sandbox boundaries and approval troubleshooting.